0% found this document useful (0 votes)
41 views94 pages

Amr 23 1014

This document reviews recent advances in applying machine learning to experimental solid mechanics. It provides an overview of common machine learning algorithms relevant to this field, such as physics-informed neural networks. The review then covers applications of machine learning in areas like fracture mechanics, biomechanics, nano- and micro-mechanics, architected materials, and 2D materials. Finally, it discusses challenges in using machine learning on multi-fidelity experimental data and quantifying prediction uncertainty, suggesting future research directions.

Uploaded by

Lexin LI
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
41 views94 pages

Amr 23 1014

This document reviews recent advances in applying machine learning to experimental solid mechanics. It provides an overview of common machine learning algorithms relevant to this field, such as physics-informed neural networks. The review then covers applications of machine learning in areas like fracture mechanics, biomechanics, nano- and micro-mechanics, architected materials, and 2D materials. Finally, it discusses challenges in using machine learning on multi-fidelity experimental data and quantifying prediction uncertainty, suggesting future research directions.

Uploaded by

Lexin LI
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 94

Applied Mechanics Reviews.

Received March 14, 2023;


Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

Recent Advances and Applications of Machine Learning in

Experimental Solid Mechanics: A Review

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


d
Hanxun Jin1,4, Enrui Zhang3, Horacio D. Espinosa1,2,*

ite
ed
1. Department of Mechanical Engineering, Northwestern University, Evanston, IL 60208

py
2. Theoretical and Applied Mechanics Program, Northwestern University, Evanston, IL

Co
60208

3. Division of Applied Mathematics, Brown University, Providence, RI 02912

ot
4. Present address: Division of Engineering and Applied Science, California Institute of
tN
Technology, Pasadena, CA 91125
rip
sc

*Corresponding Author: [email protected]


nu
Ma
ed
pt
ce
Ac

1
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

Abstract:

For many decades, experimental solid mechanics has played a crucial role in characterizing and

understanding the mechanical properties of natural and novel artificial materials. Recent advances

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


in machine learning (ML) provide new opportunities for the field, including experimental design,

d
data analysis, uncertainty quantification, and inverse problems. As the number of papers published

ite
in recent years in this emerging field is growing exponentially, it is timely to conduct a

ed
comprehensive and up-to-date review of recent ML applications in experimental solid mechanics.

py
Here, we first provide an overview of common ML algorithms and terminologies that are pertinent

Co
to this review, with emphasis placed on physics-informed and physics-based ML methods. Then,

ot
we provide thorough coverage of recent ML applications in traditional and emerging areas of
tN
experimental mechanics, including fracture mechanics, biomechanics, nano- and micro-mechanics,

architected materials, and 2D materials. Finally, we highlight some current challenges of applying
ip

ML to multi-modality and multi-fidelity experimental datasets, quantifying the uncertainty of ML


r
sc

predictions, and proposing several future research directions. This review aims to provide valuable
nu

insights into the use of ML methods and a variety of examples for researchers in solid mechanics
Ma

to integrate into their experiments.


ed
pt

Keywords:
ce

Experimental solid mechanics; Machine Learning; Fracture Mechanics; Biomechanics;


Ac

Architected Materials; Nanoindentation; 2D Materials; Physics-informed Neural Networks

(PINNs); Neural Operator

2
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

1. Introduction

Over the years, the field of experimental solid mechanics has kept evolving because of the

continuous demand to characterize and understand the mechanical properties of natural and novel

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


artificial metamaterials and structures [1, 2]. There are two primary motivations for performing

d
experiments in solid mechanics: (1) to provide experimental observations that can be used to

ite
advance universal mechanics laws; (2) to measure unknown mechanical properties of materials

ed
and structures, e.g., stiffness, strength, phase changes, inelasticity, damage, and fracture, under

py
prescribed boundary or loading conditions. Measured fields and properties guide the construction

Co
of constitutive laws and interpretation of underlying physics. Throughout the history of

ot
experimental solid mechanics, various apparatuses have been invented to measure mechanical
tN
properties, from quasi-static testing (e.g., the universal tensile testing machine [3]) to high strain-

rate testing (e.g., Kolsky bars [4] and plate impact facility [5, 6]). In the past two decades, advances
ip

in nanomechanics tools like the nanoindenter and microelectromechanical systems (MEMS) for in
r
sc

situ microscopy testing [7, 8] have enabled nanoscale characterization of advanced materials [9-
nu

11]. Likewise, various experimental measurement techniques have been developed, from local
Ma

methods, e.g., strain gauges and displacement transducers [12], to full-field methods, such as high-

resolution and high-speed imaging systems [13]. These technical innovations provide an extensive
ed

and ever-increasing amount of data collected during a single experiment. To analyze data from
pt

full-field measurement, new analysis techniques such as Moiré interferometry [14], digital
ce

image/volume correlations (DIC/DVC) [15, 16], and electronic speckle pattern interferometry [17]
Ac

have been developed to extract mechanical properties and deformation fields from experiments.

Furthermore, inverse methods have been used to extract constitutive behavior and identify

imperfections [18, 19]. Indeed, the combination of experimental mechanics with fast and robust

3
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

computational algorithms for inverse analysis has been growing in importance since it enables new

approaches to mechanical property identification, from fracture properties under extreme

conditions to anisotropic properties of biological tissues to superior mechanical properties of

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


nanoarchitected and 2D materials.

d
Recently, the concept of materials by design [20] has been advanced for the design of multi-

ite
functional architected materials [21] and 2D materials/devices [22, 23] exhibiting unprecedented

ed
performance. Such progress was possible due to the rapid development of modern experimental

py
mechanics techniques, which include fabrication processes, e.g., additive manufacturing (AM)

Co
[24], chemical vapor deposition, as well as the development of high-throughput testing

methodologies [25-27]. In such research strategy, experimental solid mechanics plays an essential

ot
role in providing valuable training and validation of experimental data for extracting physically
tN
inspired reduced-order models as well as advancing understanding of fabrication process-
ip

mechanical property relationships. Therefore, combining novel and intelligent algorithms together
r
sc

with advances in fabrication and experimental characterization methods has the potential to
nu

achieve a paradigm shift in discovering multi-functional and architected materials and structures.
Ma

Recent advances in machine learning (ML) [28], in particular, deep learning (DL) [29], offer

the opportunity to expand the field of experimental solid mechanics when combined with rapid
ed

data processing and inverse approaches. ML has played a significant role in computer science
pt

applications and technologies like computer vision [30], natural language processing [31], and
ce

self-driving cars [32]. In engineering and applied physics disciplines, ML has been widely used in
Ac

various areas of materials science, including material microstructure design [33], microscopic

imaging detection [34], and force-field development [35]. Several comprehensive reviews [36-40]

have thoroughly surveyed the potential of ML in materials science. In solid mechanics, ML has

4
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

been successfully employed in a wide range of applications, such as constructing surrogate models

for constitutive modeling [41, 42], advancing multiscale modeling [43, 44], designing architected

materials [45], extracting unknown mechanical parameters [46], or obtaining the internal material

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


information from externally measured fields [47-49]. In these applications, most ML frameworks

d
were trained on synthetic data from computational methods. Therefore, applying these ML

ite
frameworks to train and validate sparse and noisy experimental data with high fidelity and

ed
modality require cautious quantification of uncertainty [50, 51], from both experimental data and

py
ML architectures like hyperparameters, optimization method, and overparameterization. For this

Co
purpose, uncertainty quantification methods like Bayesian methods and deep ensembles can be

employed. Moreover, employing ML in the experimental design data process could not only

ot
potentially identify material properties, which could not be revealed otherwise, but also inspire
tN
experimentalists to develop new experimental techniques with metrology capable of big-data
ip

generation with high information content. With broad community interest, as reflected by the
r
sc

increasing number of publications in this field, it is timely to conduct a contemporary and thorough
nu

review of recent advances in the use of ML in experimental solid mechanics. Though this review
Ma

will be focused on experimental aspects, multidisciplinary approaches, including computational

and theoretical mechanics, as well as materials science, are needed to address engineering
ed

applications.
pt

This review is dedicated to Prof. Kyung-Suk Kim, on his 70’s birthday, to celebrate his seminal
ce

contributions to the fields of experimental and theoretical mechanics. The review is planned as
Ac

follows. In Section 2, we will briefly review some key ML algorithms that can be employed in

experimental mechanics. Then, in Section 3, we will review some recent progress of ML

applications in experimental solid mechanics, covering the broad areas of fracture mechanics,

5
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

biomechanics, nano- and micro-mechanics, mechanics of architected materials, and fracture

toughness of 2D materials. In Section 4, we will discuss how to properly select ML models for

experimental data with multi-fidelity and multi-modality. We will also discuss possible solutions

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


to estimate and reduce ML uncertainty. Finally, in Section 5, we will close our review with a

d
discussion of potential future research directions. Please note that ML has also been extensively

ite
applied to AM and full-field optical measurement, which are two other important fields in

ed
experimental solid mechanics. Since extensive state-of-the-art reviews have been conducted in

py
these two fields (see Ref. [52-54] for AM and Ref. [55] for full-field optical measurement), we

Co
will not cover these topics extensively in this review.

2. ML framework for experimental solid mechanics ot


tN
ML methods use algorithms mimicking human intelligence to perform optimization tasks for
ip

particular goals [28]. Since the early development of NNs and back-propagation algorithms [56]
r
sc

in the 1980s, the research field in ML has evolved significantly, leading to the discovery of various
nu

network architectures with distinct operational principles. Therefore, before applying these ML
Ma

algorithms to experimental solid mechanics, it is crucial to understand their architectures, working

principles, and potential limitations. Such fundamental knowledge will enable researchers to
ed

effectively understand these algorithms and utilize them appropriately in specific applications. In
pt

principle, ML methods can be divided into three main categories: supervised learning (SL),
ce

unsupervised learning (UL), and reinforcement learning (RL). In supervised learning, the
Ac

algorithm learns the mapping between the input dataset and their ground-truth labels, while in

unsupervised learning, the algorithm aims to identify patterns and features in the data without

being explicitly trained on labeled examples. In reinforced learning, an agent receives feedback in

6
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

terms of rewards or punishments for each action, and then uses these feedbacks to improve future

decision-making capabilities. In addition, there are other ML methods, such as semi-supervised

learning. In semi-supervised learning, the ML algorithm is trained with both labeled and unlabeled

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


datasets. The labeled data first train the model, and then the model is used to label the unlabeled

d
data. Examples of supervised learning algorithms include linear regression, decision trees, and

ite
neural networks (NNs). Examples of unsupervised learning algorithms include principal

ed
component analysis, K-mean clustering, and spectral clustering. Examples of RL include policy

py
gradient and Q-learning. Examples of semi-supervised learning include generative models like

Co
generative adversarial networks (GANs). For more details on the fundamentals of ML methods,

readers are referred to textbooks, see, e.g., [57] and online courses.

ot
ML methods are typically data-driven, that is, the model is trained/informed by large
tN
datasets consisting of images, texts, audio, and so on [28]. For example, the revolutionary AI
ip

chatbot software, ChatGPT [58], was trained based on a transformer language model [59] that uses
r
sc

self-attention mechanisms, allowing the model to weigh input words at different positions to
nu

predict the following words. In engineering and physics disciplines, however, many problems can
Ma

be well-defined by some underlying physical laws such as partial differential equations (PDEs).

For example, in solid mechanics, the equilibrium, compatibility, and constitutive equations (a set
ed

of PDEs) define the kinematics of continuum bodies. Moreover, the physical laws themselves
pt

provide valuable a priori temporal or spatial information, which can be integrated into the ML
ce

framework during training. To this end, the concept of physics-informed neural networks (PINNs)
Ac

[60] has been proposed by the Karniadakis group at Brown University. This framework has paved

a new pathway to solving physical-law-governed forward and inverse problems, reducing the need

to collect a large dataset. Based on the notion of data-driven and physics-informed ML methods,

7
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

we may interpolate between these two ends, formulating a spectrum of scientific ML methods that

may be developed for experimental solid mechanics. Depending on the size of the dataset we could

obtain, and how much physics is embedded in the algorithms, the scientific ML approaches can be

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


categorized into three scenarios depicted in Fig. 1: (I) physics-informed learning method [60, 61],

d
(II) physics-based data-driven method, and (III) purely data-driven method. The second scenario,

ite
i.e., the physics-based data-driven approach, is typically employed when a problem is too complex

ed
to be completely described by a set of PDEs. An example is the cohesive fracture of solids, where

py
the training dataset can be relatively easily obtained from massive computer simulations

Co
incorporating the physics of the problem, e.g., finite element analysis (FEA) or molecular

dynamics (MD).

ot
In the following subsections, we will briefly review some existing ML algorithms and
tN
introduce terminologies that are useful for applications in experimental solid mechanics. In
ip

Section 2.1, we will discuss one of the unsupervised ML methods, the clustering method. We will
r
sc

then introduce various types of NNs in Sections 2.2-2.6. Then, in Section 2.7 and 2.8, we will
nu

discuss RL and Bayesian inference. Lastly, methods used in the scientific ML community will be
Ma

discussed in Sections 2.9-2.10.


ed
pt
ce
Ac

8
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


d
ite
ed
py
Co
Figure 1 Schematics of three ML approaches based on available physics and data: (I) Physics-

ot
informed neural networks (PINNs); (II) Physics-based data-driven; and (III) Purely data-driven
tN
(Reproduced with permission from Ref. [62]. Copyright 2021 by Hanxun Jin). Figure idea from
ip

Karniadakis et al. [61].


r
sc
nu

2.1 Clustering

Clustering is an unsupervised ML technique that identifies data structure and grouping similar
Ma

datasets into several clusters. The most common clustering algorithm is the K-means method,
ed

which groups total dataset into K clusters by minimizing the total distance between the data and
pt

computed cluster centroids. When K-means algorithm fails to cluster the data due to data non-
ce

linearities or complexity, the spectral clustering algorithm [63], which uses the spectral properties
Ac

of the dataset to determine clusters, can be employed. The principle of spectral clustering is to

transform a complex dataset into a low-dimensional representation by using the spectrum

(eigenvalues) of the similarity matrix of the data. Then, the low-dimensional dataset can be

clustered using traditional clustering techniques, such as K-means clustering. In experimental solid

9
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

mechanics, spectral clustering particularly applies to acoustic signals such as acoustic emissions

from mechanical response [64]. For example, Muir et al. [65] applied the spectral clustering

technique to identify the damage mechanisms of SiC/SiC composites based on frequency

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


information of acoustic emission signals.

d
ite
2.2 Neural networks (NNs)

ed
NNs are an ML algorithm inspired by the function and structure of the human brain. They consist

py
of an input layer, interconnected layer(s), and an output layer. These layers are interconnected with

Co
nodes, which are like neurons. Through passing information by interconnected layers with

activation functions [57] (such as Sigmoid, Softmax, and Rectified Linear Unit (ReLu) functions

ot
[66]), the NN can learn a nonlinear mapping between inputs and outputs. There is a variety of NNs,
tN
including dense neural networks, convolutional neural networks (CNN), GANs, graph neural
ip

networks (GNNs), recurrent neural networks (RNNs), transformer, and autoencoder, which are
r
sc

useful in inverse problems and data generation for solid mechanics. Among these structures, the
nu

simplest architecture of NNs is the fully connected NNs, where all neurons in every layer are
Ma

connected to all neurons in the adjacent layers. In the next subsections, we will briefly overview

these NN structures and their current and potential applications in experimental solid mechanics.
ed
pt

2.3 Convolutional neural networks (CNNs)


ce

CNN is a type of DL algorithm commonly used for image classification and feature extractions
Ac

[67, 68]. The input images are processed through convolutional layers, subsequently passed

through the polling layer, and fully connected layers for feature reduction and filtering. Multiple

convolutional layers can be applied to increase the complexity of the feature extraction. CNN can

10
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

be useful for both regression and classification of image-based experimental data. For example,

CNN can be used for material property identification and microstructure characterization from

experimental images [69]. Furthermore, CNN can analyze images obtained from full-field

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


measurement techniques like interferometric and DIC data. Using CNN for feature extraction from

d
interferometric fringes can bypass the need for a conventional fringe unwrapping process while

ite
also increasing the feature extraction accuracy. For example, Jin et al. [70] employed a CNN-based

ed
DL framework to extract dynamic cohesive properties and fracture toughness of polyurea directly

py
from image-shearing interferometric fringes. Kaviani and Kolinski [71] developed a CNN-based

Co
DL framework to convert fringes from Fizeau interferometry with a low resolution into frustrated

total internal reflection images with high resolution while studying droplet impact. Another

ot
important application for CNN in experimental mechanics is to analyze DIC data. DIC is a
tN
powerful full-field measurement tool to analyze local displacement and strain distribution [15].
ip

Many advanced DIC techniques like q-factor-based DIC [72] and augmented-Lagrangian DIC [73,
r
sc

74] were developed to increase the accuracy, efficiency, and robustness of strain field calculation.
nu

Recently, Yang et al. [75] showed the pre-trained CNN-based DL algorithms from synthetic data
Ma

can accurately predict end-to-end measurement of displacement and strain fields from

experimental speckle images. CNNs were also used by Espinosa and co-workers in the study of
ed

cell morphology upon biomolecular delivery into cells using localized electroporation [76] and the
pt

localization of single cells within a population for single-cell gene editing [77].
ce
Ac

2.4 Recurrent neural networks (RNNs)

RNNs are a type of NNs that have been successfully used for processing sequential datasets such

as natural language [57]. Unlike feedforward NNs, RNNs can retain information about previous

11
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

inputs. One of the seminal RNNs models is long short-term memory (LSTM) networks [78], which

were developed to solve the vanishing gradient problem for simple RNNs. RNNs have been widely

applied in a variety of applications, such as speech recognition, machine translation, and natural

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


language processing. In solid mechanics applications, RNNs have been used in structural health

d
monitoring, such as crack path detection. For example, in Buehler’s group, LSTM models were

ite
trained based on atomistic simulations to predict the fracture patterns of crystalline solids [79] and

ed
2D materials [80, 81]. Furthermore, RNNs have been successfully applied to model solids with

py
plastic behaviors due to their capability to deal with time-dependent data. For example, Mozaffar

Co
et al. [82] applied RNNs to model path-dependent plasticity for complex microstructures.

2.5 Graph Neural Networks (GNNs) ot


tN
GNNs [83] are a type of ML method that is usually employed on graph-structured data, which can
ip

be considered as a collection of nodes and edges. The nodes represent entities, and the edges
r
sc

represent the relationships between the entities. Therefore, GNNs are well-suited for handling
nu

problems where the relationships between entities are abstract, non-sequential, and highly
Ma

interconnected. GNNs have been successfully applied to a broad range of applications, including

recommender systems [84], social networks [85], drug discovery [86], material property prediction
ed

[87], and protein nature frequency prediction [88]. In solid mechanics, GNNs have been employed
pt

to characterize and design complex mechanical materials and structures based on the graph
ce

representation of their microstructure and/or crystallography. For example, Guo and Buehler [89]
Ac

applied GNNs to design architected material through a semi-supervised approach. Xue et al. [90]

developed a GNNs-based framework to predict the nonlinear dynamics of soft mechanical

metamaterials. Hestroffer et al. [91] applied GNNs to predict mechanical properties like stiffness

12
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

and yield strength of polycrystalline materials. Thomas et al. [92] employed GNNs to represent

fatigue features in polycrystalline materials and predict high-cycle fatigue damage formation.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


2.6 Generative adversarial networks (GANs) and Conditional Generative Adversarial

d
Networks (cGANs)

ite
GANs and cGANs are ML algorithms initially used in the field of computer vision, like image

ed
generation. GANs consist of two NNs, a generator and a discriminator, trained simultaneously in

py
a game-theory-based framework to generate new synthetic datasets that mimic the original datasets

Co
[93]. First, the generator creates new data from random Gaussian noise. Then, the discriminator

evaluates if the generated data is true or false. The training finishes when a Nash Equilibrium [94]

ot
is reached, where the generator produces authentic data that the discriminator could not identify
tN
as fake. Therefore, GANs can be a promising ML algorithm for training data generation and
ip

augmentation in solid mechanics, such as generating synthetic structures of metamaterials, which


r
sc

can be used for simulations and experiments. For example, GANs were applied to generate
nu

complex architected materials, among which many of them have extreme mechanical properties
Ma

without prior knowledge [95]. Similarly, GANs were used to generate 3D microstructures from

2D sliced images, which are as authentic as real microstructures of battery electrodes [96]. While
ed

GANs have shown tremendous success in data augmentation, there are some common limitations,
pt

such as mode collapse and training instability [97], which need attention to generate high-quality
ce

synthetic data. Besides careful hyperparameter tunning, one possible solution is to introduce
Ac

physics-informed constraints during training. For instance, in the context of microstructural

generation, incorporating statistical information, such as geometry descriptors [98], as a constraint

can help ensure that the generated structures adhere to certain physical properties.

13
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

cGANs are an extension of GANs where data generation is conditioned on additional inputs or

labels [99]. Beyond its data generation capability, cGANs can also be used for image-based end-

to-end mapping and inverse problems, such as topology optimization [100]. Applying cGANs to

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


inverse problems enables the generation of solutions that are consistent with the desired properties.

d
cGANs have been applied to image-to-image transitions like image inpainting [101] or image

ite
semantic segmentation [102]. In Solid Mechanics, cGANs have been employed to inversely

ed
identify the material modulus map while employing strain/stress images [47] or predict strain and

py
stress distributions in composites [103, 104]. Furthermore, cGANs have been successfully applied

Co
to experimental data inpainting when partial experimental data is missing [70].

2.7 Reinforcement learning (RL) ot


tN
RL is a type of ML technique in which an agent learns to make the optimum decisions by
ip

interacting with its environment to achieve a specific goal [105]. During the training, the agent
r
sc

receives feedback in the form of rewards or penalties based on its actions, enabling it to adjust its
nu

time-dependent behavior to maximize the cumulative rewards. Therefore, there are three important
Ma

aspects of RL: agent, environment, and reward. The RL methods can be categorized into value-

based methods, policy-based methods, and actor-critic methods. Recent advances in deep
ed

reinforcement learning (DRL) [106] have further expanded the capability of RL for sophisticated
pt

optimization tasks. Common DRL algorithms include Deep Q-networks (DQN) [107] and deep
ce

policy gradient [108]. DRL has been widely used in game playings, such as AlphaGo [109],
Ac

autonomous robotics [110], chemical design [111], and fluid flow optimization [112]. As its

application in solid mechanics, RL has been employed in a wide range of optimization tasks, such

as materials design and structural optimization. In materials design, the geometry and material

14
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

parameters can be considered as the agent, and the desired mechanical responses, like the stress-

strain relationship, can be considered as the environment. By sampling the design space, the agent

can receive rewards when the structure reaches the desired properties. For example, Sui et al. [113]

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


applied the DQN to design biphasic materials based on desired homogenized properties. More

d
recently, Nguyen et al. [114] developed a DL method by combining GANs and RL to generate

ite
realistic 3-dimensional microstructures with user-defined structural properties.

ed
py
2.8 Bayesian inference

Co
Bayesian inference, named after Thomas Bayes, is a statistical method that allows us to quantify

the uncertainty of unknown parameters based on observed data [115]. The fundamental concept of

ot
Bayesian inference involves integrating prior knowledge about an unknown parameter with the
tN
likelihood of the observed data given that parameter to generate the posterior probability
ip

distribution. In contrast to ML methods that focus on identifying the optimal model parameters,
r
sc

i.e., the maximum likelihood estimate, Bayesian inference provides a comprehensive description
nu

of the uncertainty surrounding the parameters, allowing for robust uncertainty quantification.
Ma

However, Bayesian inference also has certain drawbacks. First, the selection of prior distribution

can be subjective and may influence the results, especially in the case of limited data. Second,
ed

computational cost can be intensive, especially for high-dimensional spaces, as it involves


pt

sampling and computing likelihood across the entire parameter space. Despite these challenges,
ce

Bayesian inference has been widely used in various fields, such as finance, environmental science,
Ac

signal processing, and healthcare. In solid mechanics, Bayesian inference has been widely used to

identify material parameters and quantify their uncertainty from various experimental data, such

15
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

as uniaxial stress-strain data [116], force-indentation depth data from nanoindentation [117], and

resonance frequency data from resonant ultrasound spectroscopy [118, 119].

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


2.9 Physics-informed Neural Networks (PINNs)

d
Since the landmark paper [60] published by the Karniadakis group in 2019 (arXiv preprint in 2017

ite
[120]), PINNs have played a significant role in scientific ML in engineering and physics

ed
disciplines [61]. The fundamental idea of PINNs is to apply a NN to approximate the solution to a

py
physical problem, where the governing physical principles (mathematically expressed by PDEs)

Co
are enforced as prior knowledge by penalizing the residuals of PDEs, similar to [121, 122]. PINNs

have been successfully employed to solve scientific problems in a widespan of engineering

ot
disciplines such as heat transfer [123], fluid dynamics [124-126], wave propagation [127], nano-
tN
optics [128], AM [129], and biomaterials [130]. Due to the injection of physical laws into the
ip

learning algorithm, PINNs require substantially less amount of data than data-driven neural
r
sc

network approaches to achieve similar predictive capability. For example, in the study on PINNs
nu

for fluid dynamics [125], by exploiting several snapshots of the concentration field of passive
Ma

scalars, PINNs are capable of predicting the velocity and pressure fields. To fulfill the same task

in a data-driven approach without integrating fluid mechanics, one may need at least hundreds of
ed

paired snapshots of concentration, velocity, and pressure fields as training data. To facilitate the
pt

usage of PINNs in the research community, Lu et al. implemented various PINN algorithms in an
ce

open-source Python library called DeepXDE [131].


Ac

As for the applications in solid mechanics, PINNs have been successfully applied to both

forward problems (i.e., solving boundary- and initial-value problems) and inverse problems (e.g.,

material characterization and defect detection). For example, Henkes et al. [132] employed PINNs

16
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

to model micromechanics for linear elastic materials. Haghighat et al. [133] applied PINNs to build

surrogate models for elastostatics and elastoplastic solids. Bastek and Kochmann [134] employed

PINNs to model the small-strain response of shell structures. Zhang et al. [48, 49] demonstrated

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


that PINNs could effectively identify the inhomogeneous material and geometry distribution under

d
plane-strain conditions. Though most of the current PINNs frameworks in solid mechanics were

ite
demonstrated using generated synthetic data as proofs of concept, these frameworks can be applied

ed
to experimental mechanics seamlessly where continuum mechanics theories apply.

py
Co
2.10 Neural Operators

NNs are not only universal approximators of continuous functions [135], but also nonlinear

ot
continuous operators [136]. Neural operators are neural network models that learn operators,
tN
which map functions to functions, such as differential operators, integral operators, and solution
ip

operators for parameterized PDEs. Learning operators is especially important in engineering and
r
sc

physics since many problems involve relationships between functions rather than between
nu

parameters. Within the scope of solid mechanics, examples of functions include displacement
Ma

fields, stress fields, load distributions, stiffness distributions, crack propagation paths, and so on.

The Karniadakis group proposed a neural operator architecture called Deep Operator Network
ed

(DeepONet) [137]. DeepONet consists of two parts: a branch net to encode the discrete input
pt

function space and a trunk net to encode the domain of the output functions. Since then, some other
ce

neural operators have been developed [138, 139]. For detailed explanations, adequate
Ac

comparations among these algorithms, and comparisons between neural operators and PINNs, the

readers can refer to the original papers and recent reviews [140, 141]. It is worth noting that while

the original versions of these neural operators are data-driven, physical principles may also be

17
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

incorporated in a similar way as PINNs, making the neural operators informed by physics in

addition to data [142, 143]. In the past few years, neural operators have been extensively applied

to diverse engineering problems. In solid mechanics applications, neural operators have been

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


successfully used for elastoplasticity [144], fracture mechanics [145], multiscale mechanics [43],

d
and biomechanics [146] [147, 148].

ite
ed
Table 1. Summary of ML applications in experimental solid mechanics

py
Areas of Detailed applications ML algorithms Selected references
experimental

Co
mechanics
Fracture Mechanics Fracture toughness NN, Decision trees Liu et al. [149, 150]
CNN Jin et al. [70]

ot
Cohesive parameters CNN Jin et al. [70]
tN
NN Su et al. [151],
Ferdousi et al. [152]
PINN-based deep- Wei et al. [153]
ip

green inversion
Crack/flaw detection PINN Zhang et al. [49]
r
sc

CNN Niu et al. [154, 155]


Crack path prediction LSTM Lew at al. [80, 81]
nu

DeepONet Goswami et al. [145]


Predict fracture Gaussian process Athanasiou et al. [156]
Ma

instability regressions
Biomechanics Human motion NN Komaris et al. [157]
CNN Eerdekens et al. [158]
ed

Constitutive parameters ResNet, CNN Holzapfel et al. [159]


pt

NN Liu et al. [160]


ce

PINN Yin et al. [130], Kamali


et al. [161]
Ac

Surrogate constitutive Thermodynamics- Masi et al. [41]


model based NN

Constitutive Linka et al. [42]


artificial NN
(CANN)

18
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

DeepONet Zhang et al. [147],


Goswami et al. [162]

Neural operator You et al. [148]

Cell manipulation and CNN Espinosa and

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


analysis coworkers

d
[76, 77, 163, 164]
Micro and Nano- Nanoindentation NN Muliana et al. [165],

ite
Mechanics Huber et al. [166, 167]
Multi-fidelity NN Lu et al. [46]

ed
(MFNN)
Bayesian method Zhang et al. [117, 168],

py
Fernandez-Zelaia et al.
[169]

Co
AFM data-driven NN Chandrashekar et al.
[170]
Microstructure CNN, U-Net Herriott et al. [171],

ot
characterizations Sepasdar et al. [172]
Random Forest Bulgarevich et al. [173]
tN
statistical algorithm
cGAN Ni et al. [47], Yang et
al. [103]
ip

Architected Verify computational CNN, Autoencoder Alderete et al. [174]


r

materials design
sc

CNN, ResNet Ma et al. [175]


nu

Training data GAN Mao et al. [95], Hsu et


generation al. [176]
Ma

GNN Guo et al. [89]


2D materials MD force-field Multi-objective Zhang et al. [177]
parameterization optimization
ed

algorithm
Fracture toughness Integrated Zhang et al. [178]
pt

experiment-
simulation
ce

framework
Ac

19
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

3 Applications of ML in experimental solid mechanics

In this section, we will review recent advancements and applications of ML in experimental solid

mechanics, covering a wide range of fields such as fracture mechanics, biomechanics, nano- and

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


micro-mechanics, architected materials, and 2D materials. Table 1 provides an overview of some

d
detailed applications in these fields, including relevant ML algorithms and representative

ite
references.

ed
3.1 ML for fracture mechanics

py
Since the landmark paper from Griffith in 1921 [179], the century-old discipline of fracture

Co
mechanics has been established in advancing a wide range of technological advancements, from

ot
airplane structural integrity to novel materials with microscale architectures to biomimicry of
tN
natural materials. From an experimental viewpoint, there are applications of material failure we

summarize here. The first deals with quantifying material intrinsic fracture properties such as
ip

fracture toughness or cohesive laws that enable the transfer of knowledge obtained from laboratory
r
sc

tests to engineering applications, e.g., designing new devices, machines, and structures with
nu

unprecedented fracture resistance. The second addresses the identification of non-visible crack-
Ma

like defects by providing information such as their location and geometries from nondestructive

evaluation (NDE) data. This enables failure and reliability analysis, carried out to prevent
ed

catastrophic failure in service.


pt

Interestingly, ML methods can estimate fracture toughness when it cannot be easily measured
ce

using traditional methods, e.g., fracture toughness testing, under quasi-static loading, based on the
Ac

American Society for Testing and Materials standard. Readers specializing in this subject are

referred to a dedicated previous review [180]. Typically, testing samples are machined into specific

geometries and dimensions, and a load is applied to the sample to propagate a crack. Then,

20
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

experimental data such as load, displacement, and crack-tip opening distance are recorded to obtain

the fracture toughness based on analytical solutions. Recently, these methods have also been

extended to measure the fracture toughness of soft materials like hydrogen [181, 182]. However,

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


such analytic solutions are not applicable when testing samples with more complex or irregular

d
geometries and/or material nonlinearities. For example, Liu et al. [149] proposed two ML

ite
approaches, decision trees and NNs, to obtain the fracture toughness for micro-fabricated ceramic

ed
cantilevers. The ML models mapped geometry descriptors of cantilevers to their fracture toughness

py
calculated from FEA. As shown in Fig. 2 (ai), when the analytical solution is not accessible due to

Co
geometry and material complexity, an ML solution trained from representative and sufficient FEA

datasets can overcome these weaknesses and provide accurate fracture toughness with a mean error

ot
of 1% (Fig. 2 (aii)). Furthermore, with knowledge extraction and transfer techniques, the fracture
tN
toughness of samples with 3D complexity can be efficiently predicted from simpler 2D simulations
ip

[150].
r
sc

Beyond the prediction of fracture toughness, ML can be employed to inversely extract


nu

cohesive law parameters (assuming its validity) from experimental measurements consisting of
Ma

load-displacement curves [151, 152] or full-field measurements. As previously articulated, the

prediction performance can be significantly improved by integrating physical governing laws like
ed

equilibrium equations into ML training. For example, Wei et al. [153] proposed a Green’s function
pt

embedded neural network to extract mixed-mode cohesive zone properties using only far-field
ce

displacement data measured from experiments. This method integrates Green’s functions as a
Ac

physical constraint, hence, reducing the amount of training data and increasing the local model

accuracy. Furthermore, ML frameworks were developed to model and characterize interfacial

behaviors. For example, Liu [183] proposed a deep material network with cohesive layers, which

21
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

enables accurate modeling of the material interface in heterogeneous materials. Wang and Sun

[184] proposed a meta-modeling method that employs deep reinforcement learning to model

constitutive behaviors of interfaces.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


Another important application of ML in fracture mechanics is the crack path prediction, given

d
the crack propagation history. Knowing the crack path is helpful in preventing catastrophic

ite
material failure by toughening the material along the path. For example, LSTM-based ML models

ed
were trained based on atomistic modeling to predict the fracture patterns of crystalline solids [79]

py
and 2D materials [80, 81]. The LSTM model is capable of learning the spatial-temporal relations

Co
from an atomic resolution of fracture, hence it is effective in predicting the crack path. In another

application, Goswami et al. [145] developed a physics-informed variational formulation of

ot
DeepONet to predict the crack path in quasi-brittle materials by mapping the initial crack
tN
configuration to the damage and displacement fields. More recently, Worthington and Chew [185]
ip

applied NNs to predict the crack path of heterogenous materials by mapping the crack process
r
sc

zone information to the possible crack growth directions based on FEA training employing a
nu

micromechanics fracture model. It is also important to mention that current ML applications for
Ma

crack path prediction are primarily trained and validated using computational datasets obtained

from FEA or MD simulations. Thus, it remains uncertain how these methods would perform when
ed

applied to real experimental data. In the future, it is necessary to obtain high-fidelity crack
pt

propagation data from advanced diagnostic imaging techniques such as in-situ computed
ce

tomography [186, 187] or in-situ electron microscopy experiments [188-192]. This approach will
Ac

enable researchers to validate and improve the current ML framework for crack path prediction.

22
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


d
ite
ed
py
Co
ot
tN
ip

Figure 2 Applications of ML in characterizing fracture cohesive properties. (a) ML solutions can


r
sc

predict accurate fracture toughness comparable to simulations when an analytical solution is not
nu

available due to sample complexity: (i) ML framework for engineering problems; (ii) NNs-based
Ma

prediction of fracture toughness (Reproduced with permission from Ref. [149]. Copyright 2020

by Elsevier). (b) A CNN-based DL algorithm can accurately determine dynamic fracture toughness
ed

and cohesive parameters under ultra-high strain rate loading: (i) DL framework for cohesive
pt

parameter inversion from dynamic big-data-generating experiments; (ii) Comparison between


ce

predicted cohesive parameters and ground-truth. (Reproduced with permission from Ref. [70].
Ac

Copyright 2022 by Elsevier).

When predicting dynamic fracture toughness under ultra-high loading rates, the conventional

experimental measurement of load-displacement is not accessible. To this end, Jin et al. [70]
23
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

proposed an ML-assisted big-data-generating experimental framework that can accurately

measure the dynamic fracture toughness and cohesive parameters of samples from plate impact

experiments. As shown in Fig. 2 (bi), a cohesive law identification experiment was developed

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


using plate impact, a target polyurea sample containing a half-plane mid-crack, and a novel spatial-

d
temporal interferometer that generates fringes associated with the sample rear surface motion

ite
history. By employing a physics-based data-driven method, using FEA, a CNN was trained to

ed
correlate the fringe images with corresponding cohesive law parameters. After the CNN was well-

py
trained, polyurea’s dynamic fracture toughness and cohesive parameters were successfully

Co
identified from the experimental fringe image (Fig. 2 (bii)). This big-data-generating experiment

framework can be easily extended to other mechanics problems under extreme conditions, such as

ot
stress wave induced phase transformations, shear localization, and others where conventional
tN
measurement methods are not applicable. Furthermore, uncertainty quantifying of dynamic
ip

fracture properties in inhomogeneous materials, e.g., composites, is crucial. Sharma et al. [193]
r
sc

developed an ML framework bridging limited experimental data from advanced experimental


nu

techniques and data-driven models like Monte Carlo simulation to quantify the uncertainty of the
Ma

dynamic fracture toughness of glass-filled epoxy composites.

Beyond fracture properties identification, ML is also effective in predicting internal crack/flaw


ed

geometry and locations from experimental measurements. For example, Zhang et al. [49] applied
pt

PINNs to identify internal cracks in linear and nonlinear solids (Fig. 3 (a)). As shown in Fig. 3
ce

(ai), the framework uses the external boundary conditions as “sensors” to inversely identify
Ac

internal cracks in the presence of deformation. More importantly, the PINNs framework directly

integrates the underlying physics, such as material compressibility and equilibrium equations, into

the loss function (Fig. 3 (aii)), hence significantly reducing the amount of data required during

24
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

training while achieving high prediction accuracy (Fig. 3 (aiii)). Using this PINNs method, the

cracks in nonlinear solids, e.g., exhibiting elastoplastic behaviors, can also be accurately identified

[194]. Moreover, by employing a similar PINNs framework, other material properties like modulus

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


distribution were inferred [48]. To predict material strength in solids with microcracks, Xu et al.

d
[195] trained an ML framework that maps the crack distribution morphology to the strength

ite
calculated from micromechanics theory. This framework can effectively predict the strength of

ed
solids with randomly distributed microcracks. ML can also be used with non-destructive

py
measurement methods to characterize internal cracks without the application of deformations. For

Co
example, Niu and Srivastava [154, 155] used FEA-trained CNN to accurately identify internal

cracks from ultrasonic measurements (Fig. 3 (bi)). Impressively, the simulation-trained CNN can

ot
make predictions of crack shape and locations, based on experimental data, with a mean average
tN
percentage error of about 5% (Fig. 3 (bii)). The accuracy of the prediction can be attributed to the
ip

fact that FEA can effectively simulate ultrasonic wave propagation processes. There is evidence
r
sc

that the method could potentially be extended to biomedical engineering applications, e.g., the
nu

identification of cancer tumors in soft breast tissue.


Ma
ed
pt
ce
Ac

25
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


d
ite
ed
py
Co
ot
tN
r ip
sc
nu
Ma

Figure 3 Applications of ML in crack/flaw detection. (a) PINN can identify internal


ed

voids/inclusions for linear and nonlinear solids: (i) General setup for geometric and material
pt

property identification; (ii) Architectures of PINNs for continuum solid mechanics. (iii) Inference
ce

of deformation patterns under different training epochs (Reproduced with permission from
Ac

Ref. [49]. Copyright 2022, The authors, published by AAAS). (b) FEA simulation-trained CNN

was used to determine crack locations and geometry in experiments: (i) Proposed CNN

architecture, (ii) CNN predicted crack property compared to ground-truth (Reproduced with

permission from Ref. [154]. Copyright 2022 by Elsevier).

26
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

3.2 ML for biomechanics

Biomechanics is an important research field addressing the mechanics of biological systems,

including organs and tissues [196]. One important research topic in biomechanics is the

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


understanding of human movement by analyzing motion data from sensors. Comprehensive and

d
massive data have been collected over the past decades, including videos of human motion

ite
kinematics, force/displacement data from wearable devices like flexible electronics, and images

ed
obtained from computed tomography and magnetic resonance imaging. Understanding these data

py
and developing applicable biomechanical models can guide the design of new devices and

Co
technologies to address body-related issues, e.g., predicting injury risk in sports and developing

advanced medical devices. Recently, ML has been widely applied for data analysis from wearable

ot
sensors [197]. For example, Komaris et al. [157] successfully trained NNs based on a public dataset
tN
of 28 professional athletes to estimate the runner’s kinetics. Eerdekens et al. [158] employed a
ip

CNN-based ML model trained from accelerometer data to understand the equine activity. Recent
r
sc

reviews [198-201] have extensively surveyed the application of ML in this area. Therefore, in this
nu

subsection, we focus on reviewing ML applications in experimental biomechanics for


Ma

characterizing and modeling biological materials such as tissues. Another important ML

application in biomechanics is cellular manipulation [202], the readers are referred to a recent
ed

review paper [163] for details.


pt

Most biological tissues, such as blood vessels and brain matter, are soft materials. Their
ce

surface can easily form multi-mode instability, i.e., Ruga morphologies [203-206], under external
Ac

loading. For example, Jin et al. [204] performed FEA and experiments to systematically

understand the surface instability and post-bifurcation phenomenon of soft matter containing

orifices, e.g., arteries, when subjected to external pressure. Furthermore, many of these materials

27
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

are anisotropic, i.e., their mechanical responses are dependent on the loading direction. Therefore,

characterization of these materials using well-planned experiments as well as identification of

constitutive models are needed to understand and predict their mechanical behaviors. In turn, the

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


information can be used in the investigation of disease and the design of artificial organs. Many

d
historical models have been developed to characterize various biomaterials, such as the neo-

ite
Hookean model [207], Ogden model [208], Fung-type model for blood vessels [209], and

ed
Holzapfel-Gasser-Ogden (HGO) model [210] for anisotropic biomaterials. However, identifying

py
these constitutive parameters from experimental data typically requires extensive nonlinear FEA

Co
and sophisticated optimization algorithms [211-213]. ML can help infer material parameters from

limited experimental data with multi-modality, such as mechanical testing and microstructure data

ot
obtained from images. For example, Liu et al. [160] developed an ML framework to identify the
tN
HGO constitutive parameters of aortic walls based on synthetic microstructural data. Kakaletsis et
ip

al. [214] compared the parameter identification accuracy from an iterative optimization framework
r
sc

and a stand-alone NN for isotropic and anisotropic biomaterials. Results suggest that replacing
nu

FEA with Gaussian process regression or NN-based metamodels could accelerate the parameter
Ma

prediction process, while replacing the entire optimization process with a stand-alone NN yielded

unsatisfactory predictions. Recently, Holzapfel et al. [159] developed a hybrid DL model based on
ed

a residual network (ResNet) and CNN to infer three unknown material parameters for a modified
pt

HGO model [215] (Fig. 4 (ai)). This ML model maps second-harmonic generation (SHG)
ce

microstructure images, representing the orientation and dispersion of collagen fibers, to their
Ac

mechanical stress-strain data for a total of 27 artery samples. By employing this hybrid ML model,

the coefficient of determination, 𝑅! , was 0.97, while conventional least square fitting gave 𝑅! =

0.676 with a much larger standard deviation (Fig. 4 (aii)). The exceptional high-accuracy

28
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

prediction achieved from a limited biomechanics dataset can be attributed to two factors. First, the

multi-modal nature of the dataset allows for the inference of material parameters beyond the

conventional stress-strain data. Second, a priori knowledge of theoretical constitutive laws also

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


contributes to reducing the necessary dataset size. As the authors suggested, expanding the

d
experimental datasets and incorporating biaxial extension experiments are required to validate

ite
these results. This research could have a transformative impact on soft tissue constitutive modeling,

ed
i.e., modeling soft tissues using prior physics laws and limited but multi-modal experimental

py
datasets.

Co
ot
tN
r ip
sc
nu
Ma
ed
pt
ce
Ac

Figure 4 Applications of ML in constitutive parameter inversion for biomaterials. (a) A hybrid

DL framework to identify unknown material parameters of arteries with high coefficient of

determination: (i) Hybrid model architecture; (ii) Predicted stress-stretch curves from standard

29
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

fitting method compared to the proposed hybrid model. (Reproduced with permission from Ref.

[159]. Copyright 2021, The authors, published by the Royal Society). (b) Non-invasive inference

of thrombus material parameters using PINNs: (i) Schematic of PINNs for solving inverse problem;

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


(ii) Prediction and ground-truth of 2D flow around a thrombus; (iii) Comparison of the inferred

d
permeability of 2D flow with the ground-truth. (Reproduced with permission from Ref.

ite
[130]. Copyright 2021 by Elsevier).

ed
py
Recent advances in PINNs provide a promising alternative in constitutive parameter

Co
identification for biomaterials by encoding the underlying physics. As shown in Fig. 4 (bi), Yin et

al. [130] employed PINNs to infer the permeability and viscoelastic modulus of the thrombus. The

ot
tN
interaction between thrombus and blood flow can be described by sets of PDEs like Cahn–Hilliard

and Navier-Stokes equations. The parameters can be accurately identified by encoding these
r ip

governing physics during the PINNs training (Fig. 4 (bii) and (biii)). Their results also demonstrated
sc

that PINNs could infer material properties from noisy data exhibiting complexity. Recently,
nu

Kamali et al. [161] implemented PINNs to accurately identify Young’s modulus and Poisson’s
Ma

ratio for heterogeneous materials like brain matter. This method has potential clinical applications,

e.g., noninvasive elastography.


ed

Recently, a data-driven computation framework for constitutive modeling was proposed


pt

by Ortiz and coworkers [216-219]. In this framework, a data-driven solver directly learns the
ce

mechanical responses of materials from experimental data, which eliminates the need for complex
Ac

empirical constitutive modeling. In this approach, NNs can be employed to build surrogate models

of biomaterials for constitutive modeling [220, 221]. For example, Linka et al. [42] developed a

constitutive artificial NN to learn the constitutive models for hyperelastic materials directly from

30
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

given stress-strain data. Masi et al. [41] introduced a thermodynamics-based NN for constitutive

modeling by coupling thermodynamics laws as constraints during training. Li and Chen [222]

developed an equilibrium-based CNN to extract local stress distribution based on DIC strain

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


measurements performed in hyperelastic materials. Wang et al. [223] developed an ML algorithm

d
based on singular value decomposition and a Gaussian process to build metamodels of constitutive

ite
laws for time-dependent and nonlinear materials. This metamodeling method can be used to

ed
determine sets of material parameters that are best fit for experimental data. In another

py
investigation, Liu et al. [224] developed a physics-informed neural network material model to

Co
characterize soft biological tissues. Their model consists of a hierarchical learning strategy by first

learning general characteristics for a class of materials and then determining parameters for each

individual case. ot
tN
Neural operators can also be employed to build data-driven surrogate models for
ip

constitutive modeling of biomaterials due to their advantage of generalizability and prediction


r
sc

efficiency for different inputs. For example, as shown in Fig. 5 (ai), Zhang et al. [147] developed
nu

a DeepONet-based model, genotype-to-biomechanical phenotype neural network (G2Φnet), to


Ma

characterize mechanical properties of soft tissues and classify their associated genotypes from

sparse and noisy experimental data. With a 2-step training process consisting of a learning stage
ed

and an inference stage with an ensemble, G2Φnet could effectively learn the constitutive models
pt

from biaxial testing data for 28 mice with 4 different genotypes with an L2 error of less than 5%
ce

Fig. 5 (aii). Interestingly, it could also identify the correct genotype. This DL framework has
Ac

important implications in biomechanics and related clinical applications, which is learning

relationships between genotype and constitutive behaviors in biological materials from limited

experimental data. Yin et al. [146] employed DeepONet to build a data-driven surrogate model

31
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

that could predict the damage progression of heterogeneous aortic walls. Goswami et al. [162]

developed a DeepONet-based surrogate model to identify pathological insults that could lead to

thoracic aortic aneurysms from a synthetic FEA database. More recently, You et al. [148]

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


employed a Fourier neural operator based method to model mechanical responses of soft tissues

d
under different loading conditions directly from experimental data (Fig. 5 (bi)). The proposed

ite
physics-guided implicit Fourier neural operator architecture is shown in Fig. 5 (bii). This method

ed
learned the material deformation model from DIC measurements and could predict the

py
displacement field under unseen loading conditions with errors smaller than those ascertained in

Co
conventional constitutive models for soft tissues (Fig. 5 (biii)).

ot
tN
r ip
sc
nu
Ma
ed
pt
ce
Ac

32
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

Figure 5 Applications of neural operator in constitutive modeling of biomaterials. (a) A

deepONet-based DL framework to infer biomechanical response and associated genotype of

tissues: (i) the DL framework; (ii) reconstructed stress-stretch relationships compared with their

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


true values (Reproduced with permission from Ref. [147]. Copyright 2022, The authors, published

d
by PLOS). (b) A neural operator model to construct the mechanical response of biological tissues

ite
from displacement data measured from DIC: (i) The biaxial experimental setup; (ii) The

ed
architecture of the physics-guided Fourier neural operator; (iii) Error comparisons of proposed

py
Fourier neural operator method and other mechanics models (Reproduced with permission from

Co
Ref. [148]. Copyright 2022 by ASME).

ot
tN
3.3 ML for Micro and Nano-mechanics

The emerging development of nanotechnology and biotechnology in recent decades continuously


ip

requires a new understanding of micro- and nano-scale material behaviors. For example, much
r
sc

research has been conducted on understanding the mechanical properties of nano and micro-pillars
nu

[225-227], thin films [228-233], nanostructured metals [27, 226, 234-237], sub-micron-sized
Ma

sensors [238], crystalline nanowires [188, 239-241], 2D materials [178, 242-246], origami [247],

nanolattice metamaterials [248-250], and copolymers with nanoscale features [70, 251-253].
ed

Conducting precise experiments to characterize these materials with small-scale features is


pt

essential to understand their properties and underlining mechanisms and developing constitutive
ce

models. There are two major steps in this process. First, experimental measurements are taken
Ac

from instrumentations with nanometer resolution. Advanced microscopes, including Scanning

Electron Microscope (SEM), Transmission Electron Microscope (TEM), and Atomic Force

Microscope (AFM), have been developed to capture images with nanometer resolution.

33
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

Furthermore, nanomechanical instruments, like nanoindenters and MEMS, have been developed

for nanomechanical characterization. Next, the properties of interest can be extracted from

measurements via inverse algorithms. However, compared to macro-scale samples, the

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


interpretation between measurable data and material properties is not straightforward because the

d
samples contain nanoscale features due to local inhomogeneity [254, 255], size effect [231, 256],

ite
or chemo-mechanical coupling [257]. For example, in instrumented indentation [258, 259],

ed
identifying the material properties from measured load and indentation depth data is nontrivial and

py
sometimes may not guarantee unique solutions when the material constitutive law is elastoplastic

Co
[260, 261] or the indentation tip has a conical shape [262]. Therefore, there is a critical demand to

identify material properties and quantify their uncertainty using nanomechanical experiments. Due

ot
to the popularity and accessibility of nanoindentation data, ML has been widely applied to such
tN
data. Hence, in this subsection, we review two methods used in interpreting nanoindentation data:
ip

the NN approach and the Bayesian-based statistical approach. We note that the methods are also
r
sc

applicable to other nanomechanical experiments, such as the membrane deflection experiments


nu

[263] and in-situ microscopy testing using MEMS technology [11], which provides direct
Ma

measurement of stresses and strains. Lastly, ML applications to other nano and micro-mechanics

features, e.g., microstructure characterization, will be briefly reviewed.


ed

For using the NNs method to identify material properties in nanoindentation, most of the
pt

training is based on FEA due to its flexibility and efficiency. Indentation, a well-defined contact
ce

mechanics problem, can be accurately simulated in either commercial FEA software or in-house
Ac

codes [264, 265]. In 2002, Muliana et al. [165] trained a neural network with hidden layers based

on 2D and 3D FEA simulations to map nonlinear material properties from simulated load-

displacement curves. They found that the trained NN can accurately predict the load-displacement

34
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

curves of materials with properties not included in the training dataset. This work demonstrated

the potential of using NNs to inversely obtain unknown material properties from experimental data.

Huber et al. [166, 167] used FEA-trained NNs to identify the Poisson’s ratio of materials exhibiting

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


plasticity with isotropic hardening, something not easily obtained before. Since then, FEA-trained

d
NNs have been widely applied as an inverse algorithm to identify material properties from

ite
nanoindentation [266-275]. However, in practice, these FEA-trained NNs could be cumbersome

ed
since they require a substantial amount of FEA training data to survey combinations of material

py
properties within specific ranges. Such a training process is generally computationally expansive,

Co
especially when the unknown parameter space is large. Furthermore, the trained NNs usually have

poor extrapolation performance, i.e., when material properties are outside the range used in the

ot
training dataset. Another issue in this approach is the lack of uncertainty quantification when
tN
identifying properties from datasets based on FEA. To overcome these challenges, Lu et al. [46]
ip

utilized a multi-fidelity NN (MFNN) [276], which trained low-fidelity FEA datasets together with
r
sc

a few high-fidelity experimental datasets together (Fig. 6(ai)). As shown in Fig. 6(aii) and (aiii),
nu

the MFNN can efficiently learn the correlations between these two datasets with different fidelities,
Ma

hence significantly increasing the identification performance while reducing the size of FEA

training datasets. The MFNN performance can be further improved by employing transfer learning
ed

when additional experimental data are obtained.


pt

The NN method can identify unknown material parameters from indentation data as
ce

explained above. However, this method could not systematically quantify the uncertainty of
Ac

identified parameters, i.e., yielding the likelihood that other possible parameter sets also minimize

the cost function. The Bayesian method, which was derived based on Bayes’ theorem [277], can

be employed to quantify uncertainty in nanoindentation since it can provide a posterior probability

35
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

for each set of parameters. For example, Fernandez-Zelaia et al. [169] utilized the Bayesian

framework to identify unknown material parameters from an FEA-trained Gaussian process

surrogate model using spherical indentation experiments. The Bayesian framework has also been

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


employed to identify unknown properties from the spherical indentation of single crystal [278],

d
plastic solids with exponential hardening laws [279], and plastically compressible solids with

ite
Deshpande–Fleck constitutive laws [168]. For the case of conical indentation, which is similar to

ed
experiments using a Berkovich indenter, the indentation force versus indentation depth (P-h) data

py
sometimes could not yield unique sets of unknown parameters. To solve this issue, as shown in

Co
Fig. 6(bi), Zhang et al. [117, 280] employed a Bayesian framework to extract plastic properties

and quantify the uncertainty from both P-h curves and surface profile datasets obtained from FE

ot
simulations (Fig. 6(bii)). Furthermore, as shown in Fig. 6(biii), the posterior probability for possible
tN
material parameters could be calculated from both noise-free and noise-contaminated datasets.
ip

Later, Zhang and Needleman [281] applied the Bayesian framework to infer power-law creep
r
sc

constitutive parameters considering both the time-dependent indentation depth data and the
nu

residual surface profile. Though this framework was developed based on synthetic data from
Ma

simulations, the parametric identification from the Bayesian approach offers valuable insights into

uncertainty quantification resulting from nanoindentation experiments performed on materials


ed

with complex constitutive behaviors.


pt

To understand the mechanical properties of soft materials such as biological tissues [282]
ce

and cells [163], one can conduct nanoindentation tests using AFM signatures based on high-
Ac

resolution force detection and cantilever tip position obtained from a four quadrant position-

sensitive photodiode [283, 284]. For example, Rajabifar et al. [285] trained a multilayer NN to

predict surface viscoelastic and adhesive properties of samples based on load-displacement curves

36
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

obtained from AFM tapping mode. They generated the training data from a rigorous contact

mechanics model, known as the enhanced Attard’s model [286]. Here also, ML can be employed

to extract the nanoscale force without complex modeling. For example, Chandrashekar et al. [170]

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


used a data-driven ML algorithm to capture the tip-sample force in dynamic AFM. The algorithm

d
was also used to successfully identify the interaction forces for two-component polymer blends.

ite
Furthermore, the material parameter identification from AFM indentation data inherently involves

ed
uncertainty, which depends on the choice of contact mechanics model. To mitigate this uncertainty

py
in modulus identification, Nguyen and Liu [287] employed five conventional supervised ML

Co
techniques (decision trees, K-nearest neighbors, linear discriminant analysis, Naïve Bayes, and

support vector machines) to classify AFM indentation curves for different materials into

ot
appropriate contact mechanics models. By choosing the appropriate contact mechanics model, the
tN
uncertainty of modulus identification can be potentially reduced.
ip

Another important ML application for nano and micro-mechanics is microstructure


r
sc

characterizations [69, 288-291]. Since the rapid advances of high-resolution imaging techniques,
nu

ML algorithms like CNNs have been increasingly applied to microstructure characterizations. For
Ma

example, ML can be employed in microstructural image segmentation to identify individual grains

or phases [292-295]. Additionally, ML can be applied to detect defects such as cracks based on
ed

microstructural images [173] as well as predict unknown material properties like yield strength
pt

[171] based on microstructure features such as grain size, orientation, and crystallographic features.
ce

Furthermore, ML can be applied to analyze evolving microstructure images to analyze time-


Ac

dependent material behaviors during fracture [296, 297], recrystallization [298], phase

transformations [299], and cell morphology [76, 77, 163, 164]. These ML applications in

37
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

microstructure characterizations could enable rapid and accurate analysis of complex material

microstructures, leading to developing next-generation materials with desired properties.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


d
ite
ed
py
Co
ot
tN
r ip
sc
nu
Ma
ed
pt

Figure 6 Applications of ML in nanoindentation. (a) DL methods including single-fidelity NNs,


ce

multi-fidelity NNs, and residual multi-fidelity NNs to identify material parameters from
Ac

instrumented indentation: (i) architectures of these NNs; (ii) MAPE as a function of training

dataset size for multi-fidelity NNs; (iii) identification of hardening exponent for two aluminum

alloys from 2D, 3D FEA simulations and three experimental data points (Reproduced with

38
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

permission from Ref. [46]. Copyright 2020, The authors, published by PNAS). (b) A Bayesian-

type statistical approach to identify material parameters and their uncertainty from conical

indentation: (i) schematics of indentation configuration with conical tips; (ii) force-indentation

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


depth curves and surface profiles for three different materials; (iii) posterior probability

d
distribution of parameters with and without noisy data (Reproduced with permission from Ref.

ite
[117]. Copyright 2019 by ASME).

ed
py
3.4 ML for architected materials

Co
Recently, the concept of materials by design has enabled us to design multi-functional materials

with unprecedented properties. Such progress can be attributed to the rapid development of

ot
computational tools for structural design and experimental techniques for the synthesis and
tN
characterization of materials [21, 300]. Among these novel materials, architected metamaterials,
ip

which combine the properties of material constitutes and architectural design, have demonstrated
r
sc

their superior mechanical properties like ultra-high specific strength [301], excellent recoverability
nu

[250, 302], and impact resilience [303]. Along this line, ML is becoming an important tool to
Ma

systematically design these novel architected metamaterials with desired properties and

functionality beyond laboratory trial-and-error [45, 304-311]. Comprehensive review papers have
ed

been published in recent years that reviewed and discussed the methodology and applications of
pt

ML in architected material design [33, 38, 312, 313]. Herein, in this subsection, we will focus on
ce

reviewing recent advances in experimental efforts in ML-enabled design of architected materials.


Ac

One significant aspect of experimental mechanics in architected materials is to verify the

predicted power of computational methods used in the inverse design. Given the formidable

challenges associated with cost and speed for the acquisition of large-scale experimental data sets,

39
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

most of the current ML frameworks in material discovery were trained based on big datasets from

reliable physically based computer simulations. Therefore, careful verifications from experiments

are necessary to assess the ML framework’s performance in producing real structures/materials

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


with desired properties. For example, as shown in Fig. 7 (a), Alderete and Pathak et al. [174] have

d
proposed an ML framework that combines the K-mean clustering methods for design space

ite
reduction and a tandem NN architecture to inversely design shape-programmable 3D Kirigami

ed
metamaterials. The tandem NN architecture was employed to circumvent the non-uniqueness issue

py
during the inverse design process (Fig. 7 (b)). The framework was trained on finite element

Co
predictions of instabilities triggering 3D out-of-plane shapes and then validated by full-field

experimental measurements using the shadow Moiré method [314] (Fig. 7 (c)). A very good

ot
agreement between ML-designed cuts and predicted out-of-plane deformation and experiments
tN
was found under mechanical actuation (stretching). Moreover, using symbolic regression, the
ip

authors could predict the onset of actuation and needed stretching to achieve specific 3D shapes.
r
sc

These programmable 3D kirigami metamaterials can be used in various engineering applications


nu

over a large range of size scales, from microscale particle trapping to macroscale solar tracking.
Ma
ed
pt
ce
Ac

40
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


d
ite
ed
py
Co
Figure 7 Applications of ML in designing shape-programable kirigami metamaterials. (a) The ML

ot
framework to inverse design kirigami metamaterials. (b) Schematics of the tandem network
tN
employed for inverse design. (c) Experimental verification of inverse design from shadow Moiré

method (Reproduced with permission from Ref. [174]. Copyright 2022, The authors, published by
r ip

Springer Nature).
sc
nu

Likewise, ML can also assist in the inverse design of complex functional soft materials.
Ma

For example, as shown in Fig. 8 (a), Ma et al. [175] developed a ResNet-based model trained with

FEA to inversely design tunable magneto-mechanical metamaterials. The ResNet model was
ed

chosen due to its capability to preserve information between shallow and deeper layers (Fig. 8 (b))
pt

[315]. The predicted structures were printed and actuated in the magnetic field and found to be in
ce

good agreement with ML predictions (Fig. 8 (c)). Moreover, GANs can also be employed to design
Ac

architected structures without prior experience. For example, Mao et al. [95] used GANs to

systematically design complex architected materials and found that some structures can reach

Hashin-Shtrikman upper bounds. Experimental verification was also conducted by the authors to

41
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

show the robustness of the framework. Furthermore, due to the stochastic deformation modes of

metamaterials like buckling under compression, uncertainty quantification is necessary to design

reliable structures. For example, Bessa et al. [316] demonstrated that a data-driven Bayesian ML

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


framework could enable the design of super compressible metamaterials (Fig. 9 (a)). Experiments

d
on multiscale 3D printed structures (both macro-scale and micro-scale) have also demonstrated

ite
the super compressibility predicted by simulations (Fig. 9 (b) and (c)).

ed
Another important feature in this field is the ability to conduct high-throughput

py
experiments to generate high-fidelity data needed for ML training. Until now, most of the ML

Co
algorithms applied to the inverse design of architected metamaterials have been trained using

computer simulations. However, such a computer data-driven approach assumes the model is

ot
accurate and readily obtainable. This may not be the case in more complex behaviors arising from
tN
material nonlinearities and rate dependencies for which accurate constitutive descriptions do not
ip

exist. A solution would be to run autonomous experiments based on a large number of additively
r
sc

manufactured structures/samples to obtain valuable representative data such as stress-strain curves


nu

or deformation patterns [317-319]. As such, there is a need to develop new ML frameworks for
Ma

the inverse design of novel metamaterials directly from large and noisy experimental datasets. For

example, Lew and Buehler [320] trained an ML framework called DeepBuckle that combines the
ed

Variational Autoencoder (VAE) model and LSTM model to quantitatively learn buckling
pt

behaviors of polymer beams from simple and limited mechanical testing on 3D printed structures.
ce
Ac

42
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


d
ite
ed
py
Co
ot
tN
r ip
sc

Figure 8 (a) Applications of ML in the inverse design of magneto-activate mechanical


nu

metamaterials. (b) The deep residual network (ResNet) architecture. (c) Comparisons between

FEA and experiments (Reproduced with permission from Ref. [175]. Copyright 2022 by ACS).
Ma
ed
pt
ce
Ac

43
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


d
ite
ed
py
Co
ot
tN
r ip
sc
nu
Ma
ed

Figure 9 (a) Data-driven Bayesian ML framework for supercompressible metamaterials design.


pt
ce

(b) Experimental validation for designed structure from fused filament fabrication using polylactic
Ac

acid. (c) Experimental validation for designed structure with micro-scale size from two-photon

lithography. The scale bar in (c) is 50 µm. (Reproduced with permission from Ref. [316].

Copyright 2019, The Authors, published by WILEY).

44
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

3.5 ML for 2D materials fracture toughness characterization

In the past decades, significant progress has been made in the synthesis of 2D materials such as

graphene, hexagonal Boron Nitride (h-BN), and transition metal dichalcogenides (TMDs), e.g.,

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


molybdenum disulfide (MoS2), including fabrication, chemical functionalization, transfer, and

d
device assembly [321-323]. Therefore, quantifying their mechanical properties, like fracture

ite
toughness, is crucial to ensure the durability and reliability of these 2D material devices. Ni et al.

ed
thoroughly reviewed the recent experimental, theoretical, and computational progress on

py
quantifying 2D materials’ fracture properties [324]. In addition, progress on ML prediction of 2D

Co
material fracture, including fracture pattern characterization, was previously discussed. Here, in

this subsection, we will focus on reviewing the most recent experimental mechanics efforts from

ot
the Espinosa group to quantify fracture toughness via in-situ high-resolution transmission electron
tN
microscopy experiments. Likewise, we will review ML-based parametrization of interatomic
ip

potentials for 2D materials and discuss the integrated experimental-computational framework


r
sc

advanced by the Espinosa group, not only to understand fracture (Fig. 10 (a)). Such ML and
nu

atomistic experimentation frameworks are essential in advancing the predictive power of atomistic
Ma

models employed in the exploration of families of 2D materials (e.g., TMDs and MXenes) in the

spirit of the materials genome initiative.


ed

The fracture toughness of monolayer graphene [325] and h-BN [245] have been measured
pt

using in-situ SEM and TEM experiments. It was found that both materials obey Griffith’s brittle
ce

fracture criterion [179]. Assisted by molecular dynamics simulations, researchers attributed their
Ac

high toughness to different intrinsic toughening mechanisms like pre-existing grain boundaries for

graphene and structural-asymmetry induced crack branching and deflection for h-BN. However,

the direct observations of atomistic information like internal defects and lattice deformation are

45
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

absent due to the resolution limits of the previous experiments. To overcome these challenges, in

our group, Zhang et al. [178] performed in-situ high-resolution transmission electron microscopy

(HRTEM) fracture experiments to investigate the fracture toughness of two TMDs, MoS2 and

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


MoSe2 (Fig. 10 (b)). The J-integral was computed from experimental stress-strain fields obtained

d
from an affine transformation (deformation gradient) using HRTEM images of atomic structures

ite
surrounding the crack tip. The experimental measurements revealed a nonlinear region near the

ed
crack tip, where bond dissociation occurs, and confirmed brittle fracture and the applicability of

py
Griffith’s fracture criterion.

Co
To verify the experimental observations, Zhang et al. also performed MD simulations with

a newly developed ML-based parametrization framework [177]. Force-field accuracy in MD

ot
simulations plays a crucial role in studying large atomic deformation as they occur near the crack
tN
tip. Although several successful parametrizations have been developed for various 2D materials
ip

[326-328], a force field that can accurately predict phase transition and fracture toughness of 2D
r
sc

materials was not reported. While ML has immense potential for force-field parameterization
nu

directly from large datasets based on first principal calculation like density functional theory (DFT)
Ma

[329, 330], a suitable methodology that can incorporate atomic configurations far from equilibrium,

essential to the prediction of fracture, is not common. Zhang et al. [177] proposed a parametrization
ed

framework trained from DFT datasets and an evolutionary multi-objective optimization algorithm
pt

(Fig. 10 (c)). This parametrization was performed iteratively and consisted of three essential steps:
ce

training, screening, and evaluation. The force field was trained to capture both near-equilibrium
Ac

properties like cohesive energy and nonequilibrium properties such as bond dissociation energy

landscapes and vacancy formation energies. Therefore, the parametrized potential can accurately

capture bond breakage during fracture. MD simulations with this potential gave similar fracture

46
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

toughness as those obtained from experimental measurements. We anticipate that by integrating

in-situ experiments with atomistic resolution and MD simulations with force-field parametrized

based on physical ML training, the fracture behaviors of other 2D materials beyond the TMDs

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


family as well as other functional crystals can be identified. Moreover, MD simulations with ML

d
parameterized force fields should accurately predict the effect of defects such as random vacancies,

ite
line vacancies, and grain boundaries, enabling the exploration of other constituents as well as

ed
thermo-mechanical properties.

py
Co
ot
tN
r ip
sc
nu
Ma
ed
pt
ce
Ac

Figure 10 Applications of ML in fracture toughness characterization of 2D materials. (a) An

integrated experiment-simulation framework to quantitively measure intrinsic fracture toughness

of 2D materials. (b) In-situ HRTEM experiments of MoSe2 compared with MD simulations

(Reproduced with permission from Ref. [178]. Copyright 2022 by National Academy of Sciences).

47
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

(c) Schematic of ML-based interatomic potential parametrization approach consisting of three

steps: training, screening, and evaluation (Reproduced with permission from Ref. [177] . Copyright

2021, The authors, published by Springer Nature).

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


d
4 Discussion

ite
4.1 Selection of an appropriate ML model for a specific experimental mechanics application

ed
Selecting an appropriate ML model for a given task in experimental mechanics can be challenging

py
since the broad range of ML algorithms and the complexity of experimental data. Here, we provide

Co
some practical guidelines on ML model selection based on the type and amount of available data,

ot
as well as the underlying physics of the problem. Before searching for a specific ML model,
tN
researchers and practitioners should ask themselves a crucial question: is ML necessary for the

problem at hand? It is important to avoid the unnecessary use of ML when conventional methods
r ip

and techniques are sufficient. For instance, using ML for constitutive parameter fitting for simple
sc

models like neo-Hookean solids from stress-strain data is not necessary. In such a case, nonlinear
nu

regression is preferred instead.


Ma

Next, it is important to properly define the types of problems that the user wishes to tackle.

Are we using ML for an inverse problem like predicting unknown material parameters from
ed

experimental data? Or is it an optimization problem like the optimal design of architected materials
pt

subject to specific constraints? Having these questions and others answered before searching for
ce

suitable ML algorithms is important. Furthermore, one should be aware of the quantity and quality
Ac

of available information before choosing an appropriate ML model. What forms of experimental

data are available (e.g., stress-strain curves, images, full-field displacement)? What is the precision

of these data? Is there any prior knowledge or simulation tool that may enrich the data? When a

48
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

large amount of experimental data can be collected, a complex ML model like NNs can be applied.

On the other hand, when there is data scarcity due to experimental constraints, the user may opt

for employing PINNs or Gaussian Processes to embed additional physics and quantify uncertainty.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


When experimental data is not adequate, possible data augmentation techniques can be

d
employed. One common approach is to interpolate among the original datasets. For image-based

ite
data, methods such as rotation, flipping, or noise injection can be employed [331]. It is crucial to

ed
exercise caution when applying these data augmentation methods, ensuring the key characteristics

py
of the original data are maintained while increasing its quantity. Furthermore, if experiments are

Co
well-defined and can be simulated efficiently using some computational tools like FEA, MD, and

DFT, we can generate synthetic data from these simulations. However, it is crucial to thoroughly

ot
examine and understand the differences and biases present in both computational and experimental
tN
datasets. We should ensure that the computational data accurately captures the experimental data's
ip

key features. During training, it is important to monitor the model's performance on both datasets
r
sc

during training to identify any overfitting to the synthetic data. Understanding the differences
nu

between the fidelity of computational and experimental data is another important aspect, which
Ma

will be discussed in the subsequent subsection.

The selection of an appropriate ML algorithm is also based on the experimental data type. If
ed

the data consists of images or video-based datasets, CNNs could be the most efficient. If the
pt

problem involves time-dependent data, RNNs and transformers could be a better choice. When the
ce

problem can be framed in terms of physical laws, PINNs should be employed. Given the speed at
Ac

which the field is evolving, literature searches should be performed to identify the most suitable

ML models for the problem at hand. Once the ML algorithms have been selected, the user can start

to build an ML pipeline using open-source ML platforms such as TensorFlow [332], PyTorch

49
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

[333], or JAX [334]. Iterative refinement of the ML model by hyperparameter-tuning based on

testing and validation datasets is necessary to reach the desired accuracy. However, it is worth

noting that over-tuning hyperparameters can potentially result in overfitting. To address this issue,

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


we can employ robust hyperparameter optimization methods such as 𝑘-fold cross-validation [335].

d
In 𝑘-fold cross-validation, the original dataset is split into 𝑘 equal folds. During each iteration,

ite
one-fold is used as the validation set, while the other 𝑘 − 1 folds are used for training. After

ed
completing all iterations, we can compute the average performance metric, such as mean squared

py
error, from all k validation sets. This method can efficiently reduce the risk of overfitting since it

Co
trains and evaluates the model on different subsets of the data. Furthermore, it is important to keep

in mind that there may be more than one possible choice of ML model for a specific experimental

ot
tN
mechanics problem. Therefore, it is practical to carefully evaluate the prediction accuracy and

efficiency of each model and select the model that is the best fit for a specific application.
r ip
sc

4.2 Integrating multi-modality and multi-fidelity experimental data into ML methods


nu

Conducting mechanical experiments, particularly those utilizing cutting-edge facilities and


Ma

techniques for extremely large/small time and length scales, can be both costly and time-

consuming. In many cases, researchers need to combine different experimental methods, hence
ed

gaining a better understanding of the mechanics problem. Furthermore, computer simulations may
pt

be employed to provide additional insight into the experiments. As a result, one may obtain
ce

experimental data with multi-modality and/or multi-fidelity. Multimodal data refers to the data on
Ac

an object comprising different forms and patterns, hence providing information from different

channels (e.g., language data in the forms of text and speech, data on the mechanical test sample

in the forms of images, and stress-strain curves). Multi-fidelity data refers to the measurement data

50
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

with different levels of accuracy (e.g., high/low-resolution images of a test sample; stress-strain

curves measured with load cells of different accuracies; data from real experiments and from

computer simulations). Typically, high-fidelity data are expensive and hence limited, while low-

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


fidelity data are cheap and plentiful.

d
To maximize the available information injected into learning algorithms, it is important to

ite
propose ML models that are capable of handling data with multi-modality [336] and multi-fidelity

ed
[276]. There have been some studies in applying such ML models to mechanics problems. For

py
example, Holzapfel et al. [159] developed an ML method that combines microstructural

Co
information and biomechanical tests. Trask et al. [337] proposed a framework that is capable of

conducting multimodal inference for lattice metamaterials, relating their lattice design, stress-

ot
strain curves, and microstructural images. Lu et al. [46] designed a multi-fidelity neural network
tN
for characterizing the mechanical properties of materials in instrumented indentation. While these
ip

studies, among others, have explored multi-modality and multi-fidelity ML methods for solid
r
sc

mechanics problems, further investigation is still needed to better integrate data with multi-
nu

modality and multi-fidelity from experiments and/or simulations to provide a deeper understanding
Ma

of the mechanics of materials and structures.


ed

4.3 Estimating and reducing the uncertainty of ML predictions


pt

Most ML applications in experimental solid mechanics provide a point estimation – that is, a single
ce

value as the best estimate. To further acquire information regarding the reliability and confidence
Ac

of such an estimation, one sometimes needs to quantify and/or reduce the uncertainty of the ML

predictions. In the context of experimental solid mechanics, there are diverse sources of

uncertainties coming from data and models related to almost every component in the research

51
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

workflow, including: (1) experimental implementation, such as the uncertainty of material

properties caused by the manufacturing of specimens, inaccurate enforcement of the experimental

setup (e.g., boundaries that are not perfectly clamped, approximate fulfillment of plane strain/stress

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


condition), representativeness and noisiness of data; (2) theoretical modeling, such as the

d
misspecification/oversimplification of constitutive models, ignoring dynamic effects, ignoring

ite
length scale effects, typical in micro- and nano-mechanics, as well as continuum assumptions,

ed
neglecting material and/or geometric nonlinearity, and stochasticity; (3) numerical modeling, such

py
as finite element discretization, inaccurate force fields in molecular dynamics simulation. On top

Co
of these three aspects of uncertainty, ML methods (especially NN-based methods) introduce a few

additional sources of uncertainty, including the choices of model architecture and hyperparameters,

ot
stochasticity in the training process, and transferability of the trained model, making the accurate
tN
quantification of total uncertainty a complex and time-consuming endeavor.
ip

Uncertainty quantification (UQ) is a discipline of science focusing on identifying,


r
sc

quantifying, and reducing uncertainties associated with models, numerical algorithms,


nu

experiments, and predicted outcomes or quantities of interest [338]. It is a broad area that has been
Ma

studied extensively and is not exclusively applicable to machine learning methods. For detailed

UQ methods and their applications in ML, readers are referred to related textbooks and review
ed

papers [339-341]. Here, we briefly review a few UQ methods that are extensively applied to
pt

estimating and reducing uncertainty in the context of ML applications in experimental solid


ce

mechanics. To quantify the uncertainty of mechanics systems, one of the most widely adopted
Ac

classes of methods is the Bayesian procedure. Built upon the well-established, century-old Bayes’

theorem [277], the Bayesian procedure seeks to infer the posterior distribution of variables based

on prior knowledge and measured data. Specific examples of methods involving the Bayesian

52
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

approach include the use of Gaussian process regression for modeling the nonlinear behavior of

solids [342], the creep behavior of concrete [343], and metamaterial design [316], as well as NNs

for crystal plasticity [344] and multiscale modeling of nanocomposite [345]. These studies, by

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


employing the Bayesian procedure, provide a distribution of the quantities of interest rather than a

d
single output. Another prevalent technique is the ensemble method, which combines a group of

ite
base (weak) models, which differ by the algorithm, hyperparameter, training data, and/or random

ed
seeds, to provide more accurate, more robust predictions. They may be applied to machine learning

py
models such as decision trees (random forests), support vector machines, and neural networks.

Co
Through this method, one may estimate the uncertainty of the prediction according to the spread

among base models as well as reducing the uncertainty by combining these models into an

ensemble for enhanced predictive capability. ot


tN
After training a machine learning model, one often needs to validate it by testing the
ip

performance of the model on unseen data. Aside from the simplest way of the train-test split of the
r
sc

dataset, techniques including 𝑘 -fold cross-validation, leave-one-out cross-validation, and


nu

bootstrapping are often employed. These techniques help to better assess the performance of the
Ma

trained model and estimate the uncertainty of predictions by efficiently utilizing the available

dataset. In addition to the foregoing methods related to uncertainty quantification, some other
ed

commonly invoked techniques and procedures include sensitivity analysis [346], Monte-Carlo
pt

simulations [347], and computation of confidence interval [348]. Note that the field of UQ has
ce

developed for decades, with numerous methods proposed and investigated. In the foregoing
Ac

context, we have only provided a very brief, non-exhaustive list of UQ methods that have been

commonly used in mechanics. Finally, we comment that there are still a lot of problems related to

the quantification of uncertainty associated with ML applications in mechanics. As we analyzed

53
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

at the beginning of this section, there are different sources of uncertainty throughout the research

workflow of experimental mechanics, many of which have not yet been well investigated. It is a

challenging yet worthwhile task to quantify the total uncertainty of the experimental mechanics

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


workflow addressing all relevant sources of uncertainty.

d
ite
5 Outlook: Future opportunities of ML in experimental solid mechanics

ed
5.1 ML for experimental mechanics under extreme conditions

py
In recent years, there has been an emerging interest in characterizing material properties under

Co
extreme conditions like high strain rate, high pressure, and high temperature. However, such

ot
experiments are often considered laborious and costly, requiring significant experiment
tN
preparation time. For example, in conventional plate impact experiments, which characterize

dynamic properties of materials under high strain rates, the experiment preparation time (e.g.,
ip

sample preparation, optical alignment, triggering circuits connection, interferometry) alone can
r
sc

take several days. Therefore, there is an increasing demand to design big-data-generating


nu

experiments where new high-throughput experimental techniques can be coupled with ML


Ma

methods to improve the efficiency of data collection and analysis. One recent study by Jin et al.

[70] demonstrated the benefits of big-data-generating experiments in the context of material


ed

dynamic fracture toughness and cohesive parameter determination. By leveraging a high-


pt

throughput optical interferometer and CNN-based ML model, the researchers were able to
ce

significantly increase the experimental efficiency, reducing the required number of experiments
Ac

by orders of magnitude. These results highlight the potential for ML to transform the field of

experimental mechanics, enabling researchers to characterize material properties more efficiently

and effectively under extreme conditions. Moving forward, it will be crucial to continue to develop

54
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

new experimental full-field measurement techniques for experiments under extreme conditions

like spatial-temporal interferometer [70] or stereo DIC [349] that can be effectively coupled with

ML to generate and analyze large volumes of high-fidelity data. By leveraging ML for extreme

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


mechanics, experimentalists will be capable of characterizing extreme material properties more

d
efficiently and accurately, enabling new insights and applications in designing next-generation

ite
materials and technologies, e.g., earthquake protective coatings used in architectural design.

ed
py
5.2 Design of intelligent architected materials with in-situ decision-making capabilities

Co
The design of materials with decision-making capability is a relatively new concept that has the

potential to revolutionize material design by offering unprecedented properties. With the help of

ot
ML, architected materials can be programmed to respond to real-time stimuli based on external
tN
input. Here, we present some potential research areas in which experimental mechanics and ML
ip

can contribute to the development of these intelligent architected materials. One such area is the
r
sc

development of new fabrication techniques to build novel material systems with high resolution.
nu

For example, new high-throughput AM techniques, such as the hydrogel infusion AM method
Ma

[350], can be employed to fabricate architected materials with complex geometries at a variety of

scales. Furthermore, bottom-up approaches such as self-assembly can be employed to


ed

spontaneously organize nanoscale constituents into ordered structures through intermolecular


pt

forces. Recently, scaling-up fabrication techniques [351], like holographic lithography, were
ce

employed to build centimeter-size samples with nanoscale features. Another area of focus is the
Ac

development of new actuation methods. Current actuation methods are mainly focused on passive

actuation techniques like mechanical or electromagnetic actuation while designing materials with

active actuation that can deform on demand according to local environmental stimuli would confer

55
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

intelligence to these materials. Designing materials with active actuation would involve the

development of ML algorithms capable of optimizing material structures and properties based on

a set of design objectives and constraints. For example, reinforcement learning algorithms could

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


be used to train materials to learn and respond to different loading scenarios, leading to enhanced

d
structural performance and durability. Moreover, ML algorithms could enable materials to make

ite
intelligent decisions in real-time based on environmental conditions, such as changes in

ed
temperature, humidity, or mechanical loads. This opens a wide range of potential applications for

py
intelligent architected materials in fields such as aerospace and robotics.

Co
6 Conclusions

ot
Recent advances in ML have revolutionized the field of experimental solid mechanics, allowing
tN
for efficient and accurate experimental design, data analysis, parametric or function identification,
ip

and inverse design. In this review paper, we highlight recent advances and applications of ML in
r
sc

experimental solid mechanics. We started by providing an overview of common ML algorithms


nu

and terminologies relevant to experimental mechanics, with a particular emphasis on physics-


Ma

informed and physics-based scientific ML methods. Then, we reviewed recent applications of ML

in traditional and emerging areas of experimental solid mechanics, including fracture mechanics,
ed

biomechanics, nano- and micro-mechanics, architected materials, and 2D materials. Furthermore,


pt

the review discussed current challenges in applying ML to problems involving data scarcity, multi-
ce

modality, and multi-fidelity experimental datasets. It also advances several future research
Ac

directions to address such challenges. It is hoped that this comprehensive and up-to-date review

will provide valuable insights for researchers and practitioners in solid mechanics who are

interested in employing ML to design and analyze their experiments. As the field continues to

56
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

evolve, it will also be essential to build bridges across disciplines, with the most obvious being

computational mechanics and materials sciences, to address challenges and opportunities.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


Acknowledgement

d
H.D.E. acknowledges the financial support from the Air Force Office of Scientific Research

ite
(AFOSR-FA9550-20-1-0258), National Science Foundation (grant CMMI-1953806), Office of

ed
Naval Research (grant N000142212133).

py
Co
Author Contribution

HJ and HDE conceived the initial idea to write this review and defined its content. HJ wrote the

ot
initial draft except for Section 4.2 and 4.3. EZ wrote the initial draft for Sections 4.2 and 4.3 and
tN
revised other sections. HDE supervised the project and revised the entire manuscript. All authors
ip

gave final approval for the publication.


r
sc
nu

References
Ma

1. Sciammarella, C.A. and F.M. Sciammarella, Experimental mechanics of solids. 2012: John
ed

Wiley & Sons.


pt

2. Kassner, M.E., et al., New directions in mechanics. Mechanics of Materials, 2005. 37(2-3):
ce

p. 231-259.
Ac

3. Davis, J.R., Tensile testing. 2004: ASM international.

4. Chen, W.W. and B. Song, Split Hopkinson (Kolsky) bar: design, testing and applications.

2010: Springer Science & Business Media.

57
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

5. Abou-Sayed, A.S., R.J. Clifton, and L. Hermann, The oblique-plate impact experiment.

Experimental Mechanics, 1976. 16(4): p. 127-132.

6. Espinosa, H.D. and S. Nemat-Nasser, Low-velocity impact testing. ASM Handbook, 2000.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


8: p. 539-559.

d
7. Espinosa, H.D., Y. Zhu, and N. Moldovan, Design and operation of a MEMS-based

ite
material testing system for nanomechanical characterization. Journal of

ed
Microelectromechanical Systems, 2007. 16(5): p. 1219-1231.

py
8. Zhu, Y. and H.D. Espinosa, An electromechanical material testing system for in situ

Co
electron microscopy and applications. Proceedings of the National Academy of Sciences

of the United States of America, 2005. 102(41): p. 14503-14508.

9. ot
Prorok, B.C., et al., Micro-and nanomechanics, in Encyclopedia of Nanoscience and
tN
Nanotechnology. 2004, Citeseer. p. 561-606.
ip

10. Haque, M., H. Espinosa, and H. Lee, MEMS for in situ testing—Handling, actuation,
r
sc

loading, and displacement measurements. MRS bulletin, 2010. 35(5): p. 375-381.


nu

11. Bhowmick, S., et al., Advanced microelectromechanical systems-based nanomechanical


Ma

testing: beyond stress and strain measurements. Mrs Bulletin, 2019. 44(6): p. 487-493.

12. Higson, G.R., Recent Advances in Strain Gauges. Journal of Scientific Instruments, 1964.
ed

41(7): p. 405-&.
pt

13. Tiwari, V., M.A. Sutton, and S.R. McNeill, Assessment of high speed imaging systems for
ce

2D and 3D deformation measurements: Methodology development and validation.


Ac

Experimental Mechanics, 2007. 47(4): p. 561-579.

14. Walker, C.A., A Historical Review of Moire Interferometry. Experimental Mechanics,

1994. 34(4): p. 281-299.

58
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

15. Chu, T.C., et al., Applications of Digital-Image-Correlation Techniques to Experimental

Mechanics. Experimental Mechanics, 1985. 25(3): p. 232-244.

16. Bay, B.K., et al., Digital volume correlation: Three-dimensional strain mapping using X-

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


ray tomography. Experimental Mechanics, 1999. 39(3): p. 217-226.

d
17. Petzing, J.N. and J.R. Tyrer, Recent developments and applications in electronic speckle

ite
pattern interferometry. Journal of Strain Analysis for Engineering Design, 1998. 33(2): p.

ed
153-169.

py
18. Doyle, J., Inverse methods in experimental mechanics. Recent Advances in Experimental

Co
Mechanics: In Honor of Isaac M. Daniel, 2002: p. 585-594.

19. Tanaka, M. and G.S. Dulikravich, Inverse problems in engineering mechanics. 1998:

Elsevier. ot
tN
20. Montemayor, L., V. Chernow, and J.R. Greer, Materials by design: Using architecture in
ip

material design to reach new property spaces. Mrs Bulletin, 2015. 40(12): p. 1122-1129.
r
sc

21. Xia, X.X., C.M. Spadaccini, and J.R. Greer, Responsive materials architected in space and
nu

time. Nature Reviews Materials, 2022. 7(9): p. 683-701.


Ma

22. Mas-Balleste, R., et al., 2D materials: to graphene and beyond. Nanoscale, 2011. 3(1): p.

20-30.
ed

23. Nathamgari, S.S.P., et al., Nonlinear mode coupling and one-to-one internal resonances in
pt

a monolayer WS2 nanoresonator. Nano letters, 2019. 19(6): p. 4052-4059.


ce

24. Greer, J.R. and J. Park, Additive Manufacturing of Nano- and Microarchitected Materials.
Ac

Nano Letters, 2018. 18(4): p. 2187-2188.

25. Stewart, C.A., et al., Accelerated discovery of oxidation resistant CoNi-base γ/γ’alloys with

high L12 solvus and low density. Materials & Design, 2020. 189: p. 108445.

59
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

26. Noh, J., et al., Unveiling new stable manganese based photoanode materials via theoretical

high-throughput screening and experiments. Chemical Communications, 2019. 55(89): p.

13418-13421.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


27. Lin, Z., et al., In-situ SEM high strain rate testing of large diameter micropillars followed

d
by TEM and EBSD postmortem analysis. Experimental Mechanics, 2021. 61: p. 739-752.

ite
28. Mitchell, T.M. and T.M. Mitchell, Machine learning. Vol. 1. 1997: McGraw-hill New

ed
York.

py
29. LeCun, Y., Y. Bengio, and G. Hinton, Deep learning. Nature, 2015. 521(7553): p. 436-

Co
444.

30. Krizhevsky, A., I. Sutskever, and G.E. Hinton, ImageNet classification with deep

ot
convolutional neural networks, in Proceedings of the 25th International Conference on
tN
Neural Information Processing Systems - Volume 1. 2012, Curran Associates Inc.: Lake
ip

Tahoe, Nevada. p. 1097–1105.


r
sc

31. Hinton, G., et al., Deep Neural Networks for Acoustic Modeling in Speech Recognition:
nu

The Shared Views of Four Research Groups. IEEE Signal Processing Magazine, 2012.
Ma

29(6): p. 82-97.

32. Ramos, S., et al. Detecting unexpected obstacles for self-driving cars: Fusing deep
ed

learning and geometric modeling. in 2017 IEEE Intelligent Vehicles Symposium (IV). 2017.
pt

IEEE.
ce

33. Guo, K., et al., Artificial intelligence and machine learning in design of mechanical
Ac

materials. Materials Horizons, 2021. 8(4): p. 1153-1172.

34. Yang, S.H., et al., Deep Learning-Assisted Quantification of Atomic Dopants and Defects

in 2D Materials. Advanced Science, 2021. 8(16).

60
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

35. Unke, O.T., et al., Machine Learning Force Fields. Chemical Reviews, 2021. 121(16): p.

10142-10186.

36. Choudhary, K., et al., Recent advances and applications of deep learning methods in

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


materials science. Npj Computational Materials, 2022. 8(1).

d
37. Mueller, T., A.G. Kusne, and R. Ramprasad, Machine Learning in Materials Science:

ite
Recent Progress and Emerging Applications. Reviews in Computational Chemistry, Vol

ed
29, 2016. 29: p. 186-273.

py
38. Butler, K.T., et al., Machine learning for molecular and materials science. Nature, 2018.

Co
559(7715): p. 547-555.

39. Wang, A.Y.T., et al., Machine Learning for Materials Scientists: An Introductory Guide

ot
toward Best Practices. Chemistry of Materials, 2020. 32(12): p. 4954-4965.
tN
40. Himanen, L., et al., Data-Driven Materials Science: Status, Challenges, and Perspectives.
ip

Advanced Science, 2019. 6(21).


r
sc

41. Masi, F., et al., Thermodynamics-based Artificial Neural Networks for constitutive
nu

modeling. Journal of the Mechanics and Physics of Solids, 2021. 147.


Ma

42. Linka, K., et al., Constitutive artificial neural networks: A fast and general approach to

predictive data-driven constitutive modeling by deep learning. Journal of Computational


ed

Physics, 2021. 429: p. 110010.


pt

43. Yin, M.L., et al., Interfacing finite elements with deep neural operators for fast multiscale
ce

modeling of mechanics problems. Computer Methods in Applied Mechanics and


Ac

Engineering, 2022. 402.

61
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

44. Alber, M., et al., Integrating machine learning and multiscale modeling—perspectives,

challenges, and opportunities in the biological, biomedical, and behavioral sciences. NPJ

digital medicine, 2019. 2(1): p. 115.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


45. Kumar, S., et al., Inverse-designed spinodoid metamaterials. Npj Computational Materials,

d
2020. 6(1).

ite
46. Lu, L., et al., Extraction of mechanical properties of materials through deep learning from

ed
instrumented indentation. Proceedings of the National Academy of Sciences of the United

py
States of America, 2020. 117(13): p. 7052-7062.

Co
47. Ni, B. and H.J. Gao, A deep learning approach to the inverse problem of modulus

identification in elasticity. Mrs Bulletin, 2021. 46(1): p. 19-25.

48. ot
Zhang, E., M. Yin, and G.E. Karniadakis Physics-Informed Neural Networks for
tN
Nonhomogeneous Material Identification in Elasticity Imaging. 2020. arXiv:2009.04525.
ip

49. Zhang, E.R., et al., Analyses of internal structures and defects in materials using physics-
r
sc

informed neural networks. Science Advances, 2022. 8(7).


nu

50. Psaros, A.F., et al., Uncertainty quantification in scientific machine learning: Methods,
Ma

metrics, and comparisons. Journal of Computational Physics, 2023: p. 111902.

51. Brodnik, N., et al., Perspective: Machine learning in experimental solid mechanics. Journal
ed

of the Mechanics and Physics of Solids, 2023. 173: p. 105231.


pt

52. Wang, C., et al., Machine learning in additive manufacturing: State-of-the-art and
ce

perspectives. Additive Manufacturing, 2020. 36: p. 101538.


Ac

53. Jin, Z., et al., Machine learning for advanced additive manufacturing. Matter, 2020. 3(5):

p. 1541-1556.

62
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

54. Qin, J., et al., Research and application of machine learning for additive manufacturing.

Additive Manufacturing, 2022: p. 102691.

55. Zuo, C., et al., Deep learning in optical metrology: a review. Light: Science & Applications,

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


2022. 11(1): p. 39.

d
56. Rumelhart, D.E., G.E. Hinton, and R.J. Williams, Learning representations by back-

ite
propagating errors. nature, 1986. 323(6088): p. 533-536.

ed
57. Goodfellow, I., Y. Bengio, and A. Courville, Deep learning. 2016: MIT press.

py
58. OpenAI, ChatGPT. 2022. p. https://openai.com/blog/chat-gpt/.

Co
59. Vaswani, A., et al., Attention is all you need. Advances in neural information processing

systems, 2017. 30.

60. ot
Raissi, M., P. Perdikaris, and G.E. Karniadakis, Physics-informed neural networks: A deep
tN
learning framework for solving forward and inverse problems involving nonlinear partial
ip

differential equations. Journal of Computational Physics, 2019. 378: p. 686-707.


r
sc

61. Karniadakis, G.E., et al., Physics-informed machine learning. Nature Reviews Physics,
nu

2021. 3(6): p. 422-440.


Ma

62. Jin, H., Big-Data-Driven Multi-Scale Experimental Study of Nanostructured Block

Copolymer’s Dynamic Toughness. 2021, Brown University.


ed

63. Von Luxburg, U., A tutorial on spectral clustering. Statistics and computing, 2007. 17: p.
pt

395-416.
ce

64. Muir, C., et al., Damage mechanism identification in composites via machine learning and
Ac

acoustic emission. npj Computational Materials, 2021. 7(1): p. 95.

63
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

65. Muir, C., et al., A machine learning framework for damage mechanism identification from

acoustic emissions in unidirectional SiC/SiC composites. npj Computational Materials,

2021. 7(1): p. 146.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


66. Agarap, A.F., Deep learning using rectified linear units (relu). arXiv preprint

d
arXiv:1803.08375, 2018.

ite
67. LeCun, Y., et al., Backpropagation applied to handwritten zip code recognition. Neural

ed
computation, 1989. 1(4): p. 541-551.

py
68. LeCun, Y., et al., Gradient-based learning applied to document recognition. Proceedings

Co
of the IEEE, 1998. 86(11): p. 2278-2324.

69. Holm, E.A., et al., Overview: Computer Vision and Machine Learning for Microstructural

ot
Characterization and Analysis. Metallurgical and Materials Transactions a-Physical
tN
Metallurgy and Materials Science, 2020. 51(12): p. 5985-5999.
ip

70. Jin, H., et al., Dynamic fracture of a bicontinuously nanostructured copolymer: A deep-
r
sc

learning analysis of big-data-generating experiment. Journal of the Mechanics and Physics


nu

of Solids, 2022. 164.


Ma

71. Kaviani, R. and J.M. Kolinski, High Resolution Interferometric Imaging of Liquid-Solid

Interfaces with HOTNNET. Experimental Mechanics, 2022.


ed

72. Landauer, A.K., et al., A q-Factor-Based Digital Image Correlation Algorithm (qDIC) for
pt

Resolving Finite Deformations with Degenerate Speckle Patterns. Experimental


ce

Mechanics, 2018. 58(5): p. 815-830.


Ac

73. Yang, J. and K. Bhattacharya, Augmented Lagrangian Digital Image Correlation.

Experimental Mechanics, 2019. 59(2): p. 187-205.

64
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

74. Yang, J. and K. Bhattacharya, Fast Adaptive Mesh Augmented Lagrangian Digital Image

Correlation. Experimental Mechanics, 2021. 61(4): p. 719-735.

75. Yang, R., et al., Deep DIC: Deep learning-based digital image correlation for end-to-end

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


displacement and strain measurement. Journal of Materials Processing Technology, 2022.

d
302.

ite
76. Patino, C.A., et al., Multiplexed high-throughput localized electroporation workflow with

ed
deep learning–based analysis for cell engineering. Science Advances, 2022. 8(29): p.

py
eabn7637.

Co
77. Mukherjee, P., et al., Deep Learning - Assisted Automated Single Cell Electroporation

ot
Platform for Effective Genetic Manipulation of Hard- to- Transfect Cells. Small, 2022.
tN
18(20): p. 2107795.
ip

78. Hochreiter, S. and J. Schmidhuber, Long short-term memory. Neural computation, 1997.
r

9(8): p. 1735-1780.
sc

79. Hsu, Y.C., C.H. Yu, and M.J. Buehler, Using Deep Learning to Predict Fracture Patterns
nu

in Crystalline Solids. Matter, 2020. 3(1): p. 197-211.


Ma

80. Lew, A.J., et al., Deep learning model to predict fracture mechanisms of graphene. Npj 2d

Materials and Applications, 2021. 5(1).


ed

81. Lew, A.J. and M.J. Buehler, A deep learning augmented genetic algorithm approach to
pt

polycrystalline 2D material fracture discovery and design. Applied Physics Reviews, 2021.
ce

8(4).
Ac

82. Mozaffar, M., et al., Deep learning predicts path-dependent plasticity. Proceedings of the

National Academy of Sciences, 2019. 116(52): p. 26414-26420.

65
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

83. Scarselli, F., et al., The graph neural network model. IEEE transactions on neural networks,

2008. 20(1): p. 61-80.

84. Wu, S., et al., Graph neural networks in recommender systems: a survey. ACM Computing

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


Surveys, 2022. 55(5): p. 1-37.

d
85. Fan, W., et al. Graph neural networks for social recommendation. in The world wide web

ite
conference. 2019.

ed
86. Xiong, Z., et al., Pushing the boundaries of molecular representation for drug discovery

py
with the graph attention mechanism. Journal of medicinal chemistry, 2019. 63(16): p.

Co
8749-8760.

87. Xie, T. and J.C. Grossman, Crystal graph convolutional neural networks for an accurate

ot
and interpretable prediction of material properties. Physical review letters, 2018. 120(14):
tN
p. 145301.
ip

88. Guo, K. and M.J. Buehler, Rapid prediction of protein natural frequencies using graph
r
sc

neural networks. Digital Discovery, 2022. 1(3): p. 277-285.


nu

89. Guo, K. and M.J. Buehler, A semi-supervised approach to architected materials design
Ma

using graph neural networks. Extreme Mechanics Letters, 2020. 41: p. 101029.

90. Xue, T., S. Adriaenssens, and S. Mao, Learning the nonlinear dynamics of mechanical
ed

metamaterials with graph networks. International Journal of Mechanical Sciences, 2023.


pt

238: p. 107835.
ce

91. Hestroffer, J.M., et al., Graph neural networks for efficient learning of mechanical
Ac

properties of polycrystals. Computational Materials Science, 2023. 217: p. 111894.

92. Thomas, A., et al., Materials fatigue prediction using graph neural networks on

microstructure representations. 2023.

66
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

93. Goodfellow, I., et al., Generative adversarial networks. Communications of the ACM,

2020. 63(11): p. 139-144.

94. Holt, C.A. and A.E. Roth, The Nash equilibrium: A perspective. Proceedings of the

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


National Academy of Sciences, 2004. 101(12): p. 3999-4002.

d
95. Mao, Y.W., Q. He, and X.H. Zhao, Designing complex architectured materials with

ite
generative adversarial networks. Science Advances, 2020. 6(17).

ed
96. Kench, S. and S.J. Cooper, Generating three-dimensional structures from a two-

py
dimensional slice with generative adversarial network-based dimensionality expansion.

Co
Nature Machine Intelligence, 2021. 3(4): p. 299-305.

97. Salimans, T., et al., Improved techniques for training gans. Advances in neural information

processing systems, 2016. 29. ot


tN
98. Cang, R., et al., Microstructure representation and reconstruction of heterogeneous
ip

materials via deep belief network for computational material design. Journal of
r
sc

Mechanical Design, 2017. 139(7): p. 071404.


nu

99. Mirza, M. and S. Osindero Conditional Generative Adversarial Nets. 2014.


Ma

arXiv:1411.1784.

100. Nie, Z., et al., Topologygan: Topology optimization using generative adversarial networks
ed

based on physical fields over the initial domain. Journal of Mechanical Design, 2021.
pt

143(3).
ce

101. Isola, P., et al. Image-to-Image Translation with Conditional Adversarial Networks. 2016.
Ac

arXiv:1611.07004.

102. Rezaei, M., et al. Conditional Adversarial Network for Semantic Segmentation of Brain

Tumor. 2017. arXiv:1708.05227.

67
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

103. Yang, Z.Z., et al., End-to-end deep learning method to predict complete strain and stress

tensors for complex hierarchical composite microstructures. Journal of the Mechanics and

Physics of Solids, 2021. 154.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


104. Yang, Z.Z., C.H. Yu, and M.J. Buehler, Deep learning model to predict complex stress and

d
strain fields in hierarchical composites. Science Advances, 2021. 7(15).

ite
105. Sutton, R.S. and A.G. Barto, Reinforcement learning: An introduction. 2018: MIT press.

ed
106. Mousavi, S.S., M. Schukat, and E. Howley. Deep reinforcement learning: an overview. in

py
Proceedings of SAI Intelligent Systems Conference (IntelliSys) 2016: Volume 2. 2018.

Co
Springer.

107. Mnih, V., et al., Human-level control through deep reinforcement learning. nature, 2015.

518(7540): p. 529-533. ot
tN
108. Lillicrap, T.P., et al., Continuous control with deep reinforcement learning. arXiv preprint
ip

arXiv:1509.02971, 2015.
r
sc

109. Silver, D., et al., Mastering the game of Go with deep neural networks and tree search.
nu

nature, 2016. 529(7587): p. 484-489.


Ma

110. Singh, B., R. Kumar, and V.P. Singh, Reinforcement learning in robotic applications: a

comprehensive survey. Artificial Intelligence Review, 2022: p. 1-46.


ed

111. Popova, M., O. Isayev, and A. Tropsha, Deep reinforcement learning for de novo drug
pt

design. Science advances, 2018. 4(7): p. eaap7885.


ce

112. Garnier, P., et al., A review on deep reinforcement learning for fluid mechanics. Computers
Ac

& Fluids, 2021. 225: p. 104973.

113. Sui, F., et al., Deep reinforcement learning for digital materials design. ACS Materials

Letters, 2021. 3(10): p. 1433-1439.

68
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

114. Nguyen, P.C., et al., Synthesizing controlled microstructures of porous media using

generative adversarial networks and reinforcement learning. Scientific reports, 2022.

12(1): p. 9034.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


115. Box, G.E. and G.C. Tiao, Bayesian inference in statistical analysis. 2011: John Wiley &

d
Sons.

ite
116. Rappel, H., et al., A tutorial on Bayesian inference to identify material parameters in solid

ed
mechanics. Archives of Computational Methods in Engineering, 2020. 27: p. 361-385.

py
117. Zhang, Y.P., J.D. Hart, and A. Needleman, Identification of Plastic Properties From

Co
Conical Indentation Using a Bayesian-Type Statistical Approach. Journal of Applied

Mechanics-Transactions of the Asme, 2019. 86(1).

118. ot
Rossin, J., et al., Bayesian inference of elastic constants and texture coefficients in
tN
additively manufactured cobalt-nickel superalloys using resonant ultrasound spectroscopy.
ip

Acta Materialia, 2021. 220: p. 117287.


r
sc

119. Rossin, J., et al., Single crystal elastic constants of additively manufactured components
nu

determined by resonant ultrasound spectroscopy. Materials Characterization, 2022. 192: p.


Ma

112244.

120. Raissi, M., P. Perdikaris, and G.E. Karniadakis, Physics informed deep learning (part i):
ed

Data-driven solutions of nonlinear partial differential equations. arXiv preprint


pt

arXiv:1711.10561, 2017.
ce

121. Lagaris, I.E., A. Likas, and D.I. Fotiadis, Artificial neural networks for solving ordinary
Ac

and partial differential equations. IEEE transactions on neural networks, 1998. 9(5): p.

987-1000.

69
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

122. Psichogios, D.C. and L.H. Ungar, A hybrid neural network-first principles approach to

process modeling. AIChE Journal, 1992. 38(10): p. 1499-1511.

123. Cai, S., et al., Physics-informed neural networks for heat transfer problems. Journal of

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


Heat Transfer, 2021. 143(6).

d
124. Cai, S., et al., Physics-informed neural networks (PINNs) for fluid mechanics: A review.

ite
Acta Mechanica Sinica, 2021. 37(12): p. 1727-1738.

ed
125. Raissi, M., A. Yazdani, and G.E. Karniadakis, Hidden fluid mechanics: Learning velocity

py
and pressure fields from flow visualizations. Science, 2020. 367(6481): p. 1026-1030.

Co
126. Mao, Z., A.D. Jagtap, and G.E. Karniadakis, Physics-informed neural networks for high-

ot
speed flows. Computer Methods in Applied Mechanics and Engineering, 2020. 360: p.
tN
112789.

127. Rasht - Behesht, M., et al., Physics - Informed Neural Networks (PINNs) for Wave
r ip

Propagation and Full Waveform Inversions. Journal of Geophysical Research: Solid Earth,
sc

2022. 127(5): p. e2021JB023120.


nu

128. Chen, Y., et al., Physics-informed neural networks for inverse problems in nano-optics and
Ma

metamaterials. Optics express, 2020. 28(8): p. 11618-11633.

129. Liao, S., et al., Hybrid thermal modeling of additive manufacturing processes using
ed

physics-informed neural networks for temperature prediction and parameter identification.


pt

Computational Mechanics, 2023: p. 1-14.


ce

130. Yin, M., et al., Non-invasive inference of thrombus material properties with physics-
Ac

informed neural networks. Computer Methods in Applied Mechanics and Engineering,

2021. 375: p. 113603.

70
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

131. Lu, L., et al., DeepXDE: A deep learning library for solving differential equations. SIAM

review, 2021. 63(1): p. 208-228.

132. Henkes, A., H. Wessels, and R. Mahnken, Physics informed neural networks for continuum

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


micromechanics. Computer Methods in Applied Mechanics and Engineering, 2022. 393:

d
p. 114790.

ite
133. Haghighat, E., et al., A physics-informed deep learning framework for inversion and

ed
surrogate modeling in solid mechanics. Computer Methods in Applied Mechanics and

py
Engineering, 2021. 379: p. 113741.

Co
134. Bastek, J.-H. and D.M. Kochmann, Physics-Informed Neural Networks for shell structures.

European Journal of Mechanics-A/Solids, 2023. 97: p. 104849.

135. ot
Hornik, K., M. Stinchcombe, and H. White, Multilayer feedforward networks are universal
tN
approximators. Neural networks, 1989. 2(5): p. 359-366.
ip

136. Chen, T. and H. Chen, Universal approximation to nonlinear operators by neural networks
r
sc

with arbitrary activation functions and its application to dynamical systems. IEEE
nu

Transactions on Neural Networks, 1995. 6(4): p. 911-917.


Ma

137. Lu, L., et al., Learning nonlinear operators via DeepONet based on the universal

approximation theorem of operators. Nature machine intelligence, 2021. 3(3): p. 218-229.


ed

138. Kovachki, N., et al., Neural operator: Learning maps between function spaces. arXiv
pt

preprint arXiv:2108.08481, 2021.


ce

139. Li, Z., et al., Fourier neural operator for parametric partial differential equations. arXiv
Ac

preprint arXiv:2010.08895, 2020.

140. Goswami, S., et al., Physics-informed deep neural operators networks. arXiv preprint

arXiv:2207.05748, 2022.

71
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

141. Lu, L., et al., A comprehensive and fair comparison of two neural operators (with practical

extensions) based on fair data. Computer Methods in Applied Mechanics and Engineering,

2022. 393: p. 114778.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


142. Li, Z., et al., Physics-informed neural operator for learning partial differential equations.

d
arXiv preprint arXiv:2111.03794, 2021.

ite
143. Wang, S., H. Wang, and P. Perdikaris, Learning the solution operator of parametric partial

ed
differential equations with physics-informed deeponets. Science advances, 2021. 7(40): p.

py
eabi8605.

Co
144. Li, Z., et al., Fourier neural operator with learned deformations for pdes on general

geometries. arXiv preprint arXiv:2207.05209, 2022.

145. ot
Goswami, S., et al., A physics-informed variational DeepONet for predicting crack path.
tN
Computer Methods in Applied Mechanics and Engineering, 2022. 391.
ip

146. Yin, M., et al., Simulating progressive intramural damage leading to aortic dissection
r
sc

using DeepONet: an operator–regression neural network. Journal of the Royal Society


nu

Interface, 2022. 19(187): p. 20210670.


Ma

147. Zhang, E., et al., G2Φnet: Relating genotype and biomechanical phenotype of tissues with

deep learning. PLOS Computational Biology, 2022. 18(10): p. e1010660.


ed

148. You, H., et al., A physics-guided neural operator learning approach to model biological
pt

tissues from digital image correlation measurements. Journal of Biomechanical


ce

Engineering, 2022. 144(12): p. 121012.


Ac

149. Liu, X., et al., A machine learning approach to fracture mechanics problems. Acta

Materialia, 2020. 190: p. 105-112.

72
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

150. Liu, X., et al., Knowledge extraction and transfer in data-driven fracture mechanics.

Proceedings of the National Academy of Sciences of the United States of America, 2021.

118(23).

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


151. Su, M., et al., Identification of the interfacial cohesive law parameters of FRP strips

d
externally bonded to concrete using machine learning techniques. Engineering Fracture

ite
Mechanics, 2021. 247.

ed
152. Ferdousi, S., et al., Characterize traction-separation relation and interfacial imperfections

py
by data-driven machine learning models. Scientific Reports, 2021. 11(1).

Co
153. Wei, C.J., et al., Deep-green inversion to extract traction-separation relations at material

interfaces. International Journal of Solids and Structures, 2022. 250.

154. ot
Niu, S.J. and V. Srivastava, Simulation trained CNN for accurate embedded crack length,
tN
location, and orientation from ultrasound measurements. International Journal of Solids
ip

and Structures, 2022. 242.


r
sc

155. Niu, S. and V. Srivastava, Ultrasound classification of interacting flaws using finite
nu

element simulations and convolutional neural network. Engineering with Computers, 2022.
Ma

38(5): p. 4653-4662.

156. Athanasiou, C.E., et al., Integrated simulation, machine learning, and experimental
ed

approach to characterizing fracture instability in indentation pillar-splitting of materials.


pt

Journal of the Mechanics and Physics of Solids, 2023. 170: p. 105092.


ce

157. Komaris, D.-S., et al., Predicting three-dimensional ground reaction forces in running by
Ac

using artificial neural networks and lower body kinematics. IEEE Access, 2019. 7: p.

156779-156786.

73
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

158. Eerdekens, A., et al., Automatic equine activity detection by convolutional neural networks

using accelerometer data. Computers and electronics in agriculture, 2020. 168: p. 105139.

159. Holzapfel, G.A., et al., Predictive constitutive modelling of arteries by deep learning.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


Journal of the Royal Society Interface, 2021. 18(182): p. 20210411.

d
160. Liu, M., L. Liang, and W. Sun, Estimation of in vivo constitutive parameters of the aortic

ite
wall using a machine learning approach. Computer methods in applied mechanics and

ed
engineering, 2019. 347: p. 201-217.

py
161. Kamali, A., M. Sarabian, and K. Laksari, Elasticity imaging using physics-informed neural

Co
networks: Spatial discovery of elastic modulus and Poisson's ratio. Acta Biomaterialia,

2023. 155: p. 400-409.

162. ot
Goswami, S., et al., Neural operator learning of heterogeneous mechanobiological insults
tN
contributing to aortic aneurysms. Journal of the Royal Society Interface, 2022. 19(193): p.
ip

20220410.
r
sc

163. Mukherjee, P., et al., Integrating Micro and Nano Technologies for Cell Engineering and
nu

Analysis: Toward the Next Generation of Cell Therapy Workflows. Acs Nano, 2022. 16(10):
Ma

p. 15653-15680.

164. Patino, C.A., et al., Deep learning and computer vision strategies for automated gene
ed

editing with a single-cell electroporation platform. SLAS TECHNOLOGY: Translating


pt

Life Sciences Innovation, 2021. 26(1): p. 26-36.


ce

165. Muliana, A., et al., Artificial neural network and finite element modeling of
Ac

nanoindentation tests. Metallurgical and Materials Transactions a-Physical Metallurgy and

Materials Science, 2002. 33(7): p. 1939-1947.

74
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

166. Huber, N., A. Konstantinidis, and C. Tsakmakis, Determination of Poisson's ratio by

spherical indentation using neural networks - Part I: Theory. Journal of Applied

Mechanics-Transactions of the Asme, 2001. 68(2): p. 218-223.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


167. Huber, N. and C. Tsakmakis, Determination of Poisson's ratio by spherical indentation

d
using neural networks - Part II: Identification method. Journal of Applied Mechanics-

ite
Transactions of the Asme, 2001. 68(2): p. 224-229.

ed
168. Zhang, Y.P. and A. Needleman, Characterization of plastically compressible solids via

py
spherical indentation. Journal of the Mechanics and Physics of Solids, 2021. 148.

Co
169. Fernandez-Zelaia, P., et al., Estimating mechanical properties from spherical indentation

using Bayesian approaches. Materials & Design, 2018. 147: p. 92-105.

170. ot
Chandrashekar, A., et al., Quantifying nanoscale forces using machine learning in dynamic
tN
atomic force microscopy. Nanoscale Advances, 2022. 4(9): p. 2134-2143.
ip

171. Herriott, C. and A.D. Spear, Predicting microstructure-dependent mechanical properties


r
sc

in additively manufactured metals with machine-and deep-learning methods.


nu

Computational Materials Science, 2020. 175: p. 109599.


Ma

172. Sepasdar, R., A. Karpatne, and M. Shakiba, A data-driven approach to full-field nonlinear

stress distribution and failure pattern prediction in composites using deep learning.
ed

Computer Methods in Applied Mechanics and Engineering, 2022. 397: p. 115126.


pt

173. Bulgarevich, D.S., et al., Pattern recognition with machine learning on optical microscopy
ce

images of typical metallurgical microstructures. Scientific reports, 2018. 8(1): p. 1-8.


Ac

174. Alderete, N.A., N. Pathak, and H.D. Espinosa, Machine learning assisted design of shape-

programmable 3D kirigami metamaterials. Npj Computational Materials, 2022. 8(1).

75
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

175. Ma, C.P., et al., Deep Learning-Accelerated Designs of Tunable Magneto-Mechanical

Metamaterials. Acs Applied Materials & Interfaces, 2022. 14(29): p. 33892-33902.

176. Hsu, T., et al., Microstructure generation via generative adversarial network for

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


heterogeneous, topologically complex 3d materials. Jom, 2021. 73: p. 90-102.

d
177. Zhang, X., et al., Multi-objective parametrization of interatomic potentials for large

ite
deformation pathways and fracture of two-dimensional materials. Npj Computational

ed
Materials, 2021. 7(1).

py
178. Zhang, X., et al., Atomistic measurement and modeling of intrinsic fracture toughness of

Co
two-dimensional materials. Proceedings of the National Academy of Sciences, 2022.

119(45): p. e2206756119.

179. ot
Griffith, A.A., VI. The phenomena of rupture and flow in solids. Philosophical transactions
tN
of the royal society of london. Series A, 1921. 221(582-593): p. 163-198.
ip

180. Zhu, X.K. and J.A. Joyce, Review of fracture toughness (G, K, J, CTOD, CTOA) testing
r
sc

and standardization. Engineering Fracture Mechanics, 2012. 85: p. 1-46.


nu

181. Zhang, E., et al., Fatigue fracture of nearly elastic hydrogels. Soft matter, 2018. 14(18): p.
Ma

3563-3571.

182. Sun, J.-Y., et al., Highly stretchable and tough hydrogels. Nature, 2012. 489(7414): p. 133-
ed

136.
pt

183. Liu, Z.L., Deep material network with cohesive layers: Multi-stage training and interfacial
ce

failure analysis. Computer Methods in Applied Mechanics and Engineering, 2020. 363.
Ac

184. Wang, K. and W.C. Sun, Meta-modeling game for deriving theory-consistent,

microstructure-based traction-separation laws via deep reinforcement learning. Computer

Methods in Applied Mechanics and Engineering, 2019. 346: p. 216-241.

76
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

185. Worthington, M. and H.B. Chew, Crack path predictions in heterogeneous media by

machine learning. Journal of the Mechanics and Physics of Solids, 2023. 172: p. 105188.

186. Chon, M.J., et al., Lamellae spatial distribution modulates fracture behavior and toughness

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


of african pangolin scales. Journal of the mechanical behavior of biomedical materials,

d
2017. 76: p. 30-37.

ite
187. García-Moreno, F., et al., Using X-ray tomoscopy to explore the dynamics of foaming metal.

ed
Nature communications, 2019. 10(1): p. 3762.

py
188. Espinosa, H.D., R.A. Bernal, and M. Minary-Jolandan, A Review of Mechanical and

Co
Electromechanical Properties of Piezoelectric Nanowires. Advanced Materials, 2012.

24(34): p. 4656-4675.

189. ot
Ramachandramoorthy, R., et al., High strain rate tensile testing of silver nanowires: rate-
tN
dependent brittle-to-ductile transition. Nano letters, 2016. 16(1): p. 255-263.
ip

190. Ramachandramoorthy, R., R. Bernal, and H.D. Espinosa, Pushing the envelope of in situ
r
sc

transmission electron microscopy. ACS nano, 2015. 9(5): p. 4675-4685.


nu

191. Bernal, R.A., R. Ramachandramoorthy, and H.D. Espinosa, Double-tilt in situ TEM holder
Ma

with multiple electrical contacts and its application in MEMS-based mechanical testing of

nanomaterials. Ultramicroscopy, 2015. 156: p. 23-28.


ed

192. Bernal, R.A., et al., In situ electron microscopy four - point electromechanical
pt

characterization of freestanding metallic and semiconducting nanowires. small, 2014.


ce

10(4): p. 725-733.
Ac

193. Sharma, A., T. Mukhopadhyay, and V. Kushvaha, Experimental data-driven uncertainty

quantification for the dynamic fracture toughness of particulate polymer composites.

Engineering Fracture Mechanics, 2022. 273.

77
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

194. Niu, S., et al., Modeling finite-strain plasticity using physics informed neural network and

assessment of the network performance. Journal of the Mechanics and Physics of Solids,

2022: p. 105177.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


195. Xu, B.W., et al., Deep learning method for predicting the strengths of microcracked brittle

d
materials. Engineering Fracture Mechanics, 2022. 271.

ite
196. Knudson, D.V. and D. Knudson, Fundamentals of biomechanics. Vol. 183. 2007: Springer.

ed
197. Zhang, S., et al., Deep learning in human activity recognition with wearable sensors: A

py
review on advances. Sensors, 2022. 22(4): p. 1476.

Co
198. Halilaj, E., et al., Machine learning in human movement biomechanics: Best practices,

common pitfalls, and new opportunities. Journal of biomechanics, 2018. 81: p. 1-11.

199. ot
Phellan, R., et al., Real-time biomechanics using the finite element method and machine
tN
learning: Review and perspective. Medical Physics, 2021. 48(1): p. 7-18.
ip

200. Low, W.S., et al., A review of machine learning network in human motion biomechanics.
r
sc

Journal of Grid Computing, 2022. 20(1): p. 4.


nu

201. Mouloodi, S., et al., What can artificial intelligence and machine learning tell us? A review
Ma

of applications to equine biomechanical research. Journal of the Mechanical Behavior of

Biomedical Materials, 2021. 123: p. 104728.


ed

202. Gerbin, K.A., et al., Cell states beyond transcriptomics: integrating structural organization
pt

and gene expression in hiPSC-derived cardiomyocytes. Cell Systems, 2021. 12(6): p. 670-
ce

687. e10.
Ac

203. Diab, M., et al., Ruga mechanics of creasing: from instantaneous to setback creases.

Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences,

2013. 469(2157): p. 20120753.

78
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

204. Jin, H., A.K. Landauer, and K.-S. Kim, Ruga mechanics of soft-orifice closure under

external pressure. Proceedings of the Royal Society A, 2021. 477(2249): p. 20210238.

205. Zhao, R., et al., The primary bilayer ruga-phase diagram I: Localizations in ruga evolution.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


Extreme Mechanics Letters, 2015. 4: p. 76-82.

d
206. Zhao, R., M. Diab, and K.-S. Kim, The primary bilayer ruga-phase diagram II:

ite
Irreversibility in ruga evolution. Journal of Applied Mechanics, 2016. 83(9): p. 091004.

ed
207. Treloar, L., Stresses and birefringence in rubber subjected to general homogeneous strain.

py
Proceedings of the Physical Society, 1948. 60(2): p. 135.

Co
208. Ogden, R.W., Large deformation isotropic elasticity–on the correlation of theory and

experiment for incompressible rubberlike solids. Proceedings of the Royal Society of

ot
London. A. Mathematical and Physical Sciences, 1972. 326(1567): p. 565-584.
tN
209. Fung, Y.-c., Biomechanics: motion, flow, stress, and growth. 2013: Springer Science &
ip

Business Media.
r
sc

210. Holzapfel, G.A., T.C. Gasser, and R.W. Ogden, A new constitutive framework for arterial
nu

wall mechanics and a comparative study of material models. Journal of elasticity and the
Ma

physical science of solids, 2000. 61: p. 1-48.

211. Shi, L., et al., Anisotropic material characterization of human cervix tissue based on
ed

indentation and inverse finite element analysis. Journal of biomechanical engineering,


pt

2019. 141(9).
ce

212. Kakaletsis, S., et al., Right ventricular myocardial mechanics: Multi-modal deformation,
Ac

microstructure, modeling, and comparison to the left ventricle. Acta biomaterialia, 2021.

123: p. 154-166.

79
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

213. Sugerman, G.P., et al., A whole blood thrombus mimic: constitutive behavior under simple

shear. journal of the mechanical behavior of biomedical materials, 2021. 115: p. 104216.

214. Kakaletsis, S., E. Lejeune, and M.K. Rausch, Can machine learning accelerate soft

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


material parameter identification from complex mechanical test data? Biomechanics and

d
Modeling in Mechanobiology, 2022: p. 1-14.

ite
215. Holzapfel, G.A., et al., Modelling non-symmetric collagen fibre dispersion in arterial walls.

ed
Journal of the royal society interface, 2015. 12(106): p. 20150188.

py
216. Kirchdoerfer, T. and M. Ortiz, Data-driven computational mechanics. Computer Methods

Co
in Applied Mechanics and Engineering, 2016. 304: p. 81-101.

217. Eggersmann, R., et al., Model-free data-driven inelasticity. Computer Methods in Applied

Mechanics and Engineering, 2019. 350: p. 81-99. ot


tN
218. Stainier, L., A. Leygue, and M. Ortiz, Model-free data-driven methods in mechanics:
ip

material data identification and solvers. Computational Mechanics, 2019. 64(2): p. 381-
r
sc

393.
nu

219. Prume, E., S. Reese, and M. Ortiz, Model-free Data-Driven inference in computational
Ma

mechanics. Computer Methods in Applied Mechanics and Engineering, 2023. 403: p.

115704.
ed

220. Tac, V., et al., Data-driven modeling of the mechanical behavior of anisotropic soft
pt

biological tissue. Engineering with Computers, 2022. 38(5): p. 4167-4182.


ce

221. He, Q., et al., Manifold learning based data-driven modeling for soft biological tissues.
Ac

Journal of biomechanics, 2021. 117: p. 110124.

80
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

222. Li, L. and C. Chen, Equilibrium-based convolution neural networks for constitutive

modeling of hyperelastic materials. Journal of the Mechanics and Physics of Solids, 2022.

164: p. 104931.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


223. Wang, J., et al., Metamodeling of constitutive model using Gaussian process machine

d
learning. Journal of the Mechanics and Physics of Solids, 2021. 154: p. 104532.

ite
224. Liu, M., L. Liang, and W. Sun, A generic physics-informed neural network-based

ed
constitutive model for soft biological tissues. Computer methods in applied mechanics and

py
engineering, 2020. 372: p. 113402.

Co
225. Chen, M., A.S. Sologubenko, and J.M. Wheeler, Exploring defect behavior and size effects

in micron-scale germanium from cryogenic to elevated temperatures. Matter, 2023.

226. ot
Greer, J.R. and J.T.M. De Hosson, Plasticity in small-sized metallic systems: Intrinsic
tN
versus extrinsic size effect. Progress in Materials Science, 2011. 56(6): p. 654-724.
ip

227. Gu, X.W., et al., Size-dependent deformation of nanocrystalline Pt nanopillars. Nano


r
sc

letters, 2012. 12(12): p. 6385-6392.


nu

228. Rao, Z.X., et al., Determination of Stresses in Incrementally Deposited Films From Wafer-
Ma

Curvature Measurements. Journal of Applied Mechanics-Transactions of the Asme, 2020.

87(10).
ed

229. Chason, E., et al., Origin of compressive residual stress in polycrystalline thin films.
pt

Physical Review Letters, 2002. 88(15).


ce

230. Espinosa, H., et al., Mechanical properties of ultrananocrystalline diamond thin films for
Ac

MEMS applications. MRS Online Proceedings Library (OPL), 2002. 741.

81
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

231. Espinosa, H., B. Prorok, and B. Peng, Plasticity size effects in free-standing submicron

polycrystalline FCC films subjected to pure tension. Journal of the Mechanics and Physics

of Solids, 2004. 52(3): p. 667-689.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


232. Pugno, N., B. Peng, and H. Espinosa, Predictions of strength in MEMS components with

d
defects––a novel experimental–theoretical approach. International journal of solids and

ite
structures, 2005. 42(2): p. 647-661.

ed
233. Freund, L.B. and S. Suresh, Thin film materials: stress, defect formation and surface

py
evolution. 2004: Cambridge university press.

Co
234. Greer, J.R., W.C. Oliver, and W.D. Nix, Size dependence in mechanical properties of gold

at the micron scale in the absence of strain gradients (vol 53, pg 1821, 2005). Acta

Materialia, 2006. 54(6): p. 1705-1705. ot


tN
235. Zhao, H.Z., et al., Fracture and toughening mechanisms in nanotwinned and nanolayered
ip

materials. Mrs Bulletin, 2022. 47(8): p. 839-847.


r
sc

236. Jin, H., J. Zhou, and Y. Chen, Grain size gradient and length scale effect on mechanical
nu

behaviors of surface nanocrystalline metals. Materials Science and Engineering: A, 2018.


Ma

725: p. 1-7.

237. Li, X., et al., Mechanical properties and deformation mechanisms of gradient
ed

nanostructured metals and alloys. Nature Reviews Materials, 2020. 5(9): p. 706-723.
pt

238. Lee, B., S. Roh, and J. Park, Current status of micro- and nano-structured optical fiber
ce

sensors. Optical Fiber Technology, 2009. 15(3): p. 209-221.


Ac

239. Ramachandramoorthy, R., et al., Reliability of single crystal silver nanowire-based systems:

stress assisted instabilities. ACS nano, 2017. 11(5): p. 4768-4776.

82
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

240. Bernal, R.A., et al., Intrinsic Bauschinger effect and recoverable plasticity in pentatwinned

silver nanowires tested in tension. Nano letters, 2015. 15(1): p. 139-146.

241. Espinosa, H.D., R.A. Bernal, and T. Filleter, In-Situ TEM Electromechanical Testing of

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


Nanowires and Nanotubes. Nano and Cell Mechanics: Fundamentals and Frontiers, 2013:

d
p. 191-226.

ite
242. Akinwande, D., et al., A review on mechanics and mechanical properties of 2D materials-

ed
Graphene and beyond. Extreme Mechanics Letters, 2017. 13: p. 42-77.

py
243. Wei, X., et al., Recoverable slippage mechanism in multilayer graphene leads to

Co
repeatable energy dissipation. ACS nano, 2016. 10(2): p. 1820-1828.

ot
244. Soler-Crespo, R.A., et al., Atomically thin polymer layer enhances toughness of graphene
tN
oxide monolayers. Matter, 2019. 1(2): p. 369-388.

245. Yang, Y.C., et al., Intrinsic toughening and stable crack propagation in hexagonal boron
ip

nitride. Nature, 2021. 594(7861): p. 57-+.


r
sc

246. Choi, J.Y., et al., Atomistic mechanisms of adhesion and shear strength in graphene oxide-
nu

polymer interfaces. Journal of the Mechanics and Physics of Solids, 2021. 156: p. 104578.
Ma

247. Lin, Z., et al., Folding at the microscale: Enabling multifunctional 3D origami-architected

metamaterials. Small, 2020. 16(35): p. 2002229.


ed

248. Bauer, J., et al., Nanolattices: An Emerging Class of Mechanical Metamaterials. Advanced
pt

Materials, 2017. 29(40).


ce

249. Vyatskikh, A., et al., Additive manufacturing of 3D nano-architected metals. Nature


Ac

communications, 2018. 9(1): p. 593.

250. Meza, L.R., S. Das, and J.R. Greer, Strong, lightweight, and recoverable three-dimensional

ceramic nanolattices. Science, 2014. 345(6202): p. 1322-1326.

83
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

251. Bang, J., et al., Block Copolymer Nanolithography: Translation of Molecular Level

Control to Nanoscale Patterns. Advanced Materials, 2009. 21(47): p. 4769-4792.

252. Jin, H., et al., Understanding the Nanoscale Deformation Mechanisms of Polyurea from In

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


Situ AFM Tensile Experiments, in Challenges in Mechanics of Time Dependent Materials,

d
Mechanics of Biological Systems and Materials & Micro-and Nanomechanics, Volume 2.

ite
2022, Springer. p. 45-51.

ed
253. Kim, K.-S., et al. Dynamic fracture-toughness testing of a hierarchically nano-structured

py
solid. in Fracture, Fatigue, Failure and Damage Evolution, Volume 3: Proceedings of the

Co
2020 Annual Conference on Experimental and Applied Mechanics. 2021. Springer.

254. Xia, S.M., et al., Strength characterization of Al/Si interfaces: A hybrid method of

ot
nanoindentation and finite element analysis. Acta Materialia, 2009. 57(3): p. 695-707.
tN
255. Shen, Y.L., et al., Correlation between tensile and indentation behavior of particle-
ip

reinforced metal matrix composites: An experimental and numerical study. Acta Materialia,
r
sc

2001. 49(16): p. 3219-3229.


nu

256. Nix, W.D. and H.J. Gao, Indentation size effects in crystalline materials: A law for strain
Ma

gradient plasticity. Journal of the Mechanics and Physics of Solids, 1998. 46(3): p. 411-

425.
ed

257. Papakyriakou, M., M. Lu, and S.M. Xia, Nanoindentation Size Effects in Lithiated and
pt

Sodiated Battery Electrode Materials. Journal of Applied Mechanics-Transactions of the


ce

Asme, 2022. 89(7).


Ac

258. Oliver, W.C. and G.M. Pharr, An Improved Technique for Determining Hardness and

Elastic-Modulus Using Load and Displacement Sensing Indentation Experiments. Journal

of Materials Research, 1992. 7(6): p. 1564-1583.

84
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

259. Oliver, W.C. and G.M. Pharr, Measurement of hardness and elastic modulus by

instrumented indentation: Advances in understanding and refinements to methodology.

Journal of Materials Research, 2004. 19(1): p. 3-20.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


260. Alkorta, J., J.M. Martinez-Esnaola, and J.G. Sevillano, Absence of one-to-one

d
correspondence between elastoplastic properties and sharp-indentation load-penetration

ite
data (vol 20, pg 432, 2005). Journal of Materials Research, 2005. 20(5): p. 1369-1369.

ed
261. Chen, X., et al., On the uniqueness of measuring elastoplastic properties from indentation:

py
The indistinguishable mystical materials. Journal of the Mechanics and Physics of Solids,

Co
2007. 55(8): p. 1618-1660.

262. Cheng, Y.T. and C.M. Cheng, Can stress-strain relationships be obtained from indentation

ot
curves using conical and pyramidal indenters? Journal of Materials Research, 1999. 14(9):
tN
p. 3493-3496.
ip

263. Espinosa, H., B. Prorok, and M. Fischer, A methodology for determining mechanical
r
sc

properties of freestanding thin films and MEMS materials. Journal of the Mechanics and
nu

Physics of Solids, 2003. 51(1): p. 47-67.


Ma

264. Needleman, A., V. Tvergaard, and E. Van der Giessen, Indentation of elastically soft and

plastically compressible solids. Acta Mechanica Sinica, 2015. 31(4): p. 473-480.


ed

265. Bower, A.F., et al., Indentation of a Power Law Creeping Solid. Proceedings of the Royal
pt

Society-Mathematical and Physical Sciences, 1993. 441(1911): p. 97-124.


ce

266. Lee, H., et al., An Investigation of Nanomechanical Properties of Materials using


Ac

Nanoindentation and Artificial Neural Network. Scientific Reports, 2019. 9.

85
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

267. Konstantopoulos, G., E.P. Koumoulos, and C.A. Charitidis, Classification of mechanism

of reinforcement in the fiber-matrix interface: Application of Machine Learning on

nanoindentation data. Materials & Design, 2020. 192.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


268. Weng, J., et al., A machine learning based approach for determining the stress-strain

d
relation of grey cast iron from nanoindentation. Mechanics of Materials, 2020. 148.

ite
269. Haj-Ali, R., et al., Nonlinear constitutive models from nanoindentation tests using artificial

ed
neural networks. International Journal of Plasticity, 2008. 24(3): p. 371-396.

py
270. Han, G., K.P. Marimuthu, and H. Lee, Evaluation of thin film material properties using a

Co
deep nanoindentation and ANN. Materials & Design, 2022. 221.

271. Jeong, K., et al., Deep learning-based indentation plastometry in anisotropic materials.

International Journal of Plasticity, 2022. 157. ot


tN
272. Li, H., et al., Identification of material properties using nanoindentation and surrogate
ip

modeling. International Journal of Solids and Structures, 2016. 81: p. 151-159.


r
sc

273. Kim, Y., et al., Novel deep learning approach for practical applications of indentation.
nu

Materials Today Advances, 2022. 13.


Ma

274. Xia, J., et al., Artificial Neural Networks for Predicting Plastic Anisotropy of Sheet Metals

Based on Indentation Test. Materials, 2022. 15(5).


ed

275. Tyulyukovskiy, E. and N. Huber, Identification of viscoplastic material parameters from


pt

spherical indentation data: Part I. Neural networks. Journal of Materials Research, 2006.
ce

21(3): p. 664-676.
Ac

276. Meng, X.H. and G.E. Karniadakis, A composite neural network that learns from multi-

fidelity data: Application to function approximation and inverse PDE problems. Journal of

Computational Physics, 2020. 401.

86
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

277. Bayes, T., LII. An essay towards solving a problem in the doctrine of chances. By the late

Rev. Mr. Bayes, FRS communicated by Mr. Price, in a letter to John Canton, AMFR S.

Philosophical transactions of the Royal Society of London, 1763(53): p. 370-418.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


278. Castillo, A. and S.R. Kalidindi, A Bayesian Framework for the Estimation of the Single

d
Crystal Elastic Parameters From Spherical Indentation Stress-Strain Measurements.

ite
Frontiers in Materials, 2019. 6.

ed
279. Wang, M.Z. and J.J. Wu, Identification of plastic properties of metal materials using

py
spherical indentation experiment and Bayesian model updating approach. International

Co
Journal of Mechanical Sciences, 2019. 151: p. 733-745.

280. Zhang, Y. and A. Needleman, Influence of assumed strain hardening relation on plastic

ot
stress-strain response identification from conical indentation. Journal of Engineering
tN
Materials and Technology, 2020. 142(3).
ip

281. Zhang, Y.P. and A. Needleman, On the identification of power-law creep parameters from
r
sc

conical indentation. Proceedings of the Royal Society a-Mathematical Physical and


nu

Engineering Sciences, 2021. 477(2252).


Ma

282. Asgari, M., et al., Revealing Layer-Specific Ultrastructure and Nanomechanics of Fibrillar

Collagen in Human Aorta via Atomic Force Microscopy Testing: Implications on Tissue
ed

Mechanics at Macroscopic Scale. Advanced Nanobiomed Research, 2022. 2(5).


pt

283. Broussard, J.A., et al., The desmoplakin–intermediate filament linkage regulates cell
ce

mechanics. Molecular biology of the cell, 2017. 28(23): p. 3156-3164.


Ac

284. Broussard, J.A., et al., Scaling up single-cell mechanics to multicellular tissues–the role of

the intermediate filament–desmosome network. Journal of cell science, 2020. 133(6): p.

jcs228031.

87
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

285. Rajabifar, B., et al., Machine Learning Approach to Characterize the Adhesive and

Mechanical Properties of Soft Polymers Using PeakForce Tapping AFM. Macromolecules,

2022. 55(19): p. 8731-8740.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


286. Attard, P., Interaction and deformation of elastic bodies: origin of adhesion hysteresis. The

d
Journal of Physical Chemistry B, 2000. 104(45): p. 10635-10641.

ite
287. Nguyen, L.T.P. and B.H. Liu, Machine learning approach for reducing uncertainty in AFM

ed
nanomechanical measurements through selection of appropriate contact model. European

py
Journal of Mechanics a-Solids, 2022. 94.

Co
288. Chan, H., et al., Machine learning enabled autonomous microstructural characterization

in 3D samples. npj Computational Materials, 2020. 6(1): p. 1.

289. ot
Bostanabad, R., et al., Computational microstructure characterization and reconstruction:
tN
Review of the state-of-the-art techniques. Progress in Materials Science, 2018. 95: p. 1-41.
ip

290. Ge, M., et al., Deep learning analysis on microscopic imaging in materials science.
r
sc

Materials Today Nano, 2020. 11: p. 100087.


nu

291. Chowdhury, A., et al., Image driven machine learning methods for microstructure
Ma

recognition. Computational Materials Science, 2016. 123: p. 176-187.

292. DeCost, B.L., et al., High throughput quantitative metallography for complex
ed

microstructures using deep learning: A case study in ultrahigh carbon steel. Microscopy
pt

and Microanalysis, 2019. 25(1): p. 21-29.


ce

293. Chen, Z. and S. Daly, Deformation twin identification in magnesium through clustering
Ac

and computer vision. Materials Science and Engineering: A, 2018. 736: p. 61-75.

294. Durmaz, A.R., et al., A deep learning approach for complex microstructure inference.

Nature communications, 2021. 12(1): p. 6272.

88
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

295. Stuckner, J., B. Harder, and T.M. Smith, Microstructure segmentation with deep learning

encoders pre-trained on a large microscopy dataset. npj Computational Materials, 2022.

8(1): p. 200.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


296. Liu, Q., et al., Machine-learning assisted laser powder bed fusion process optimization for

d
AlSi10Mg: New microstructure description indices and fracture mechanisms. Acta

ite
Materialia, 2020. 201: p. 316-328.

ed
297. Müller, A., et al., Machine learning classifiers for surface crack detection in fracture

py
experiments. International Journal of Mechanical Sciences, 2021. 209: p. 106698.

Co
298. Hashemi, S. and S.R. Kalidindi, A machine learning framework for the temporal evolution

of microstructure during static recrystallization of polycrystalline materials simulated by

ot
cellular automaton. Computational Materials Science, 2021. 188: p. 110132.
tN
299. Li, Y. and S. Li, Deep learning based phase transformation model for the prediction of
ip

microstructure and mechanical properties of hot-stamped parts. International Journal of


r
sc

Mechanical Sciences, 2022. 220: p. 107134.


nu

300. Alberi, K., et al., The 2019 materials by design roadmap. Journal of Physics D-Applied
Ma

Physics, 2019. 52(1).

301. Zhang, X., et al., Lightweight, flaw-tolerant, and ultrastrong nanoarchitected carbon.
ed

Proceedings of the National Academy of Sciences of the United States of America, 2019.
pt

116(14): p. 6665-6672.
ce

302. Portela, C.M., et al., Extreme mechanical resilience of self-assembled nanolabyrinthine


Ac

materials. Proceedings of the National Academy of Sciences of the United States of

America, 2020. 117(11): p. 5686-5693.

89
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

303. Portela, C.M., et al., Supersonic impact resilience of nanoarchitected carbon. Nature

Materials, 2021. 20(11): p. 1491-+.

304. Gu, G.X., et al., Bioinspired hierarchical composite design using machine learning:

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


simulation, additive manufacturing, and experiment. Materials Horizons, 2018. 5(5): p.

d
939-945.

ite
305. Bastek, J.H., et al., Inverting the structure-property map of truss metamaterials by deep

ed
learning. Proceedings of the National Academy of Sciences of the United States of

py
America, 2022. 119(1).

Co
306. Wu, L.L., et al., A machine learning -based method to design modular metamaterials.

Extreme Mechanics Letters, 2020. 36.

307. ot
Wang, L.W., et al., Deep generative modeling for mechanistic-based learning and design
tN
of metamaterial systems. Computer Methods in Applied Mechanics and Engineering, 2020.
ip

372.
r
sc

308. Cecen, A., et al., Material structure-property linkages using three-dimensional


nu

convolutional neural networks. Acta Materialia, 2018. 146: p. 76-84.


Ma

309. Deng, B.L., et al., Inverse Design of Mechanical Metamaterials with Target Nonlinear

Response via a Neural Accelerated Evolution Strategy. Advanced Materials, 2022. 34(41).
ed

310. Cheng, X., et al., Programming 3D curved mesosurfaces using microlattice designs.
pt

Science, 2023. 379(6638): p. 1225-1232.


ce

311. Bai, Y., et al., A dynamically reprogrammable surface with self-evolving shape morphing.
Ac

Nature, 2022. 609(7928): p. 701-708.

312. Muhammad, J. Kennedy, and C.W. Lim, Machine learning and deep learning in phononic

crystals and metamaterials-A review. Materials Today Communications, 2022. 33.

90
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

313. Wang, Y.Z., et al., Inverse design of shell-based mechanical metamaterial with customized

loading curves based on machine learning and genetic algorithm. Computer Methods in

Applied Mechanics and Engineering, 2022. 401.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


314. Alderete, N.A., et al., Programmable 3D structures via Kirigami engineering and

d
controlled stretching. Extreme Mechanics Letters, 2021. 43.

ite
315. He, K., et al. Deep residual learning for image recognition. in Proceedings of the IEEE

ed
conference on computer vision and pattern recognition. 2016.

py
316. Bessa, M.A., P. Glowacki, and M. Houlder, Bayesian Machine Learning in Metamaterial

Co
Design: Fragile Becomes Supercompressible. Advanced Materials, 2019. 31(48).

317. Gongora, A.E., et al., A Bayesian experimental autonomous researcher for mechanical

ot
design. Science advances, 2020. 6(15): p. eaaz1708.
tN
318. Gongora, A.E., et al., Using simulation to accelerate autonomous experimentation: A case
ip

study using mechanics. Iscience, 2021. 24(4): p. 102262.


r
sc

319. Stach, E., et al., Autonomous experimentation systems for materials development: A
nu

community perspective. Matter, 2021. 4(9): p. 2702-2726.


Ma

320. Lew, A.J. and M.J. Buehler, DeepBuckle: Extracting physical behavior directly from

empirical observation for a material agnostic approach to analyze and predict buckling.
ed

Journal of the Mechanics and Physics of Solids, 2022. 164.


pt

321. Akinwande, D., et al., Graphene and two-dimensional materials for silicon technology.
ce

Nature, 2019. 573(7775): p. 507-518.


Ac

322. Geim, A.K., Graphene: Status and Prospects. Science, 2009. 324(5934): p. 1530-1534.

91
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

323. Dong, S.Y., et al., Facile fabrication of 2D material multilayers and vdW heterostructures

with multimodal microscopy and AFM characterization. Materials Today, 2022. 52: p. 31-

42.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


324. Ni, B., et al., Fracture at the two-dimensional limit. Mrs Bulletin, 2022. 47(8): p. 848-862.

d
325. Zhang, P., et al., Fracture toughness of graphene. Nature Communications, 2014. 5.

ite
326. Jiang, J.W., H.S. Park, and T. Rabczuk, Molecular dynamics simulations of single-layer

ed
molybdenum disulphide (MoS2): Stillinger-Weber parametrization, mechanical properties,

py
and thermal conductivity. Journal of Applied Physics, 2013. 114(6).

Co
327. Ostadhossein, A., et al., ReaxFF Reactive Force-Field Study of Molybdenum Disulfide

(MoS2). Journal of Physical Chemistry Letters, 2017. 8(3): p. 631-640.

328. ot
Wen, M.J., et al., A force-matching Stillinger-Weber potential for MoS2: Parameterization
tN
and Fisher information theory based sensitivity analysis. Journal of Applied Physics, 2017.
ip

122(24).
r
sc

329. Botu, V., et al., Machine Learning Force Fields: Construction, Validation, and Outlook.
nu

Journal of Physical Chemistry C, 2017. 121(1): p. 511-522.


Ma

330. Li, Y., et al., Machine learning force field parameters from ab initio data. Journal of

chemical theory and computation, 2017. 13(9): p. 4492-4503.


ed

331. Shorten, C. and T.M. Khoshgoftaar, A survey on image data augmentation for deep
pt

learning. Journal of big data, 2019. 6(1): p. 1-48.


ce

332. Abadi, M., et al., Tensorflow: Large-scale machine learning on heterogeneous distributed
Ac

systems. arXiv preprint arXiv:1603.04467, 2016.

333. Paszke, A., et al., Pytorch: An imperative style, high-performance deep learning library.

Advances in neural information processing systems, 2019. 32.

92
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

334. Bradbury, J., et al., JAX: composable transformations of Python+ NumPy programs. 2018.

335. Kohavi, R. A study of cross-validation and bootstrap for accuracy estimation and model

selection. in Ijcai. 1995. Montreal, Canada.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


336. Ngiam, J., et al. Multimodal deep learning. in Proceedings of the 28th international

d
conference on machine learning (ICML-11). 2011.

ite
337. Trask, N., et al., Unsupervised physics-informed disentanglement of multimodal data for

ed
high-throughput scientific discovery. arXiv preprint arXiv:2202.03242, 2022.

py
338. Smith, R.C., Uncertainty quantification: theory, implementation, and applications. Vol. 12.

Co
2013: Siam.

339. Abdar, M., et al., A review of uncertainty quantification in deep learning: Techniques,

ot
applications and challenges. Information Fusion, 2021. 76: p. 243-297.
tN
340. Soize, C., Uncertainty quantification. 2017: Springer.
ip

341. Sullivan, T.J., Introduction to uncertainty quantification. Vol. 63. 2015: Springer.
r
sc

342. Cicci, L., et al., Uncertainty quantification for nonlinear solid mechanics using reduced
nu

order models with Gaussian process regression. arXiv preprint arXiv:2302.08216, 2023.
Ma

343. Liang, M., et al., Interpretable Ensemble-Machine-Learning models for predicting creep

behavior of concrete. Cement and Concrete Composites, 2022. 125: p. 104295.


ed

344. de Oca Zapiain, D.M., et al., Predicting plastic anisotropy using crystal plasticity and
pt

Bayesian neural network surrogate models. Materials Science and Engineering: A, 2022.
ce

833: p. 142472.
Ac

345. Pyrialakos, S., et al., A neural network-aided Bayesian identification framework for

multiscale modeling of nanocomposites. Computer Methods in Applied Mechanics and

Engineering, 2021. 384: p. 113937.

93
Applied Mechanics Reviews. Received March 14, 2023;
Accepted manuscript posted July 17, 2023. doi:10.1115/1.4062966
Copyright (c) 2023 by ASME

346. Nguyen, M.S.T. and S.-E. Kim, A hybrid machine learning approach in prediction and

uncertainty quantification of ultimate compressive strength of RCFST columns.

Construction and Building Materials, 2021. 302: p. 124208.

Downloaded from http://asmedigitalcollection.asme.org/appliedmechanicsreviews/article-pdf/doi/10.1115/1.4062966/7024790/amr-23-1014.pdf by Bibliotheek Tu Delft user on 18 July 2023


347. Huang, T., et al., Microstructure-guided deep material network for rapid nonlinear

d
material modeling and uncertainty quantification. Computer Methods in Applied

ite
Mechanics and Engineering, 2022. 398: p. 115197.

ed
348. Huang, D.Z., et al., Learning constitutive relations from indirect observations using deep

py
neural networks. Journal of Computational Physics, 2020. 416: p. 109491.

Co
349. Ravindran, S., et al., Three dimensional full-field velocity measurements in shock

compression experiments using stereo digital image correlation. arXiv preprint

arXiv:2210.12568, 2022. ot
tN
350. Saccone, M.A., et al., Additive manufacturing of micro-architected metals via hydrogel
ip

infusion. Nature, 2022: p. 1-2.


r
sc

351. Kagias, M., et al., Metasurface-Enabled Holographic Lithography for Impact-Absorbing


nu

Nano-Architected Sheets. Advanced Materials: p. 2209153.


Ma
ed
pt
ce
Ac

94

You might also like