Received: 30 September 2020 Revised: 5 November 2020 Accepted: 16 November 2020 Published on: 4 March 2021

DOI: 10.1002/gamm.202100003

O R I G I N A L PA P E R

Machine learning for material characterization with an

application for predicting mechanical properties

Anke Stoll1 Peter Benner2

1
Data-Mining and Machine Learning,
Fraunhofer Institute for Machine Tools Abstract
and Forming Technology IWU, Chemnitz, Currently, the growth of material data from experiments and simulations is
Germany
expanding beyond processable amounts. This makes the development of new
2
Computational Methods in Systems and
data-driven methods for the discovery of patterns among multiple length-
Control Theory, The Max Planck Institute
for Dynamics of Complex Technical scales and time-scales and structure-property relationships essential. These
Systems, Magdeburg, Germany data-driven approaches show enormous promise within materials science. The
Correspondence
following review covers machine learning (ML) applications for metallic mate-
Anke Stoll, Reichenhainer Straße 88, rial characterization. Many parameters associated with the processing and the
09126 Chemnitz, Germany. Email: structure of materials affect the properties and the performance of manufac-
[email protected]
tured components. Thus, this study is an attempt to investigate the usefulness of
ML methods for material property prediction. Material characteristics such as
strength, toughness, hardness, brittleness, or ductility are relevant to categorize
a material or component according to their quality. In industry, material tests
like tensile tests, compression tests, or creep tests are often time consuming and
expensive to perform. Therefore, the application of ML approaches is considered
helpful for an easier generation of material property information. This study
also gives an application of ML methods on small punch test (SPT) data for the
determination of the property ultimate tensile strength for various materials. A
strong correlation between SPT data and tensile test data was found which ulti-
mately allows to replace more costly tests by simple and fast tests in combination
with ML.

KEYWORDS
machine learning, material characterization, small punch test, tensile properties, ultimate tensile
strength

1 I N T RO DU CT ION

The field of materials science relies on experiments and simulation-based models as tools for material characterization [7].
Material properties, such as their structure and behavior, are critical to the potential application of the material of inter-
est. More recently, the data generated by such experiments and simulations have created various opportunities for the
application of data-driven methods. In addition to, for example, the experimental trial and error approach or a physical
metallurgy approach, machine learning (ML) methods for property prediction and material design have attracted a lot of
attention in recent years, see for example, [33, 109, 162].
This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium,
provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.
© 2021 The Authors. GAMM - Mitteilungen published by Wiley-VCH GmbH.

GAMM - Mitteilungen. 2021;44:e202100003. wileyonlinelibrary.com/journal/gamm 1 of 21

https://doi.org/10.1002/gamm.202100003
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2 of 21 STOLL and BENNER

While experimental investigations (the so-called first paradigm of materials science) have been carried out since the
stone and copper age, scientists of the 16th century started to describe physical relations by equations (second paradigm).
Thus, analytical equations became a central instrument of theoretical physics which were able to complement the empir-
ical and experimental sciences. The 1950s marked the beginning of computational materials science and simulations, the
third paradigm. Within this framework, computer experiments and simulations became possible, with the correspond-
ing results being analyzed and interpreted like measured ones. It had to be recognized that many properties of materials
cannot be described by a closed mathematical form as they are determined by several multilevel, intricate theoretical
concepts. With the help of large amounts of data, hidden correlations, reflected in terms of structure and patterns in the
data can be discovered that are not normally visible in small data sets. Thus, the fourth paradigm, data-driven science, of
materials research was born [7, 34].
However, it is not only an advantage to have a large data volume but it can also be a challenge to cope with tremendous
amounts of data. Today, data are indeed more and more easily acquired and stored, due to huge progresses in sensors and
ways to collect data on one side, and in storage devices on the other side. Nowadays, there is no hesitation in many domains
in acquiring very large amounts of data without knowing in advance if they will be analyzed and how. The spectacular
increase in the amount of data is not only found in the number of samples collected for example over time, but also in the
number of attributes, or characteristics, that are simultaneously measured on a process. Data are gathered into vectors
whose dimension corresponds to the number of simultaneous measurements on the process. Growing dimensions result
in high dimensional data, as each sample can be represented as a point or vector in a high-dimensional space. Working
with high-dimensional data means working with data that are embedded in high-dimensional spaces [159]. The curse
of dimensionality is the expression of all phenomena that appear with high-dimensional data, and that have most often
unfortunate consequences on the behavior and performances of learning algorithms.
Contrary to the curse of dimensionality, databases in materials science are often limited in size due to expensive and
time consuming data acquisition via experiments or simulations [136]. Then the insufficient data size for the training of
a ML model compromises the learning success and suitable new approaches for small datasets have to be found.
This work contains a literature survey which covers an overview of ML for materials science and specifically for metal-
lic material characterization. As the measurement of such parameters is often expensive and time consuming obtained
via experiments, alternative basic tests, such as the small punch test (SPT) can be an option if it can be shown that the
same material property information can be extracted.
There is a wide range of ML approaches based on SPT data which will be presented. Furthermore, in Section 3 an
example is described which uses ML for the prediction of tensile properties of an insert-material-type based on SPT
data. The objective of this study is to investigate whether it is possible to find a ML model which predicts/determines
the tensile properties of a material from SPT data [11]. Section sec4 concludes this paper by giving an outlook on further
research perspectives.

2 STAT E O F TH E ART

2.1 Overview—ML for materials science

With ML, given enough data and a data-driven algorithm for rule discovery, a computer is able to determine physical
laws which lead to the given data without human input [19, 68]. Traditional computational approaches use the computer
for the employment of a hard-coded algorithm provided by a human expert. By contrast, ML approaches learn the rules
that underlie a dataset by assessing a portion of that data and building a model to make predictions [19]. However, the
human still needs to choose suitable ML models which supposedly represent the data well and do manual (sub-)tasks in
preprocessing and feature generation.
The existence of large amounts of data makes the use of ML models possible and enables data-driven knowledge to be
obtained and patterns to be discovered. On the other hand, big data and their high dimensionality lead to difficult compu-
tational and statistical challenges, such as scalability and memory shortage, noise accumulation, interference correlation,
incidental endogeneity, and measurement errors [40].
Materials science is an interesting field of application for big data methods and ML approaches which is begin-
ning to show enormous promise. Four primary elements are critical in materials science and engineering: processing,
structure, properties, and performance [110, 115]. There is no general agreement, however, on how these elements
are interconnected. ML methods can be applied to the so-called process-structure-property-performance chain for
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
STOLL and BENNER 3 of 21

learning more about the intrinsic interrelations of these components. One main goal is the enabling, acceleration, and
simplification of the discovery and development of novel materials based on the convergence of high-performance com-
puting, automation, and ML [27]. Another aim of using such approaches in the field of materials science is to achieve
high-throughput identification and quantification of essential material properties [15].
Besides experimentally obtained datasets, numerous studies draw required information from simulation-based data
mining. Altogether, it is shown that experiment- and simulation-based data mining in combination with ML tools pro-
vide exceptional opportunities to enable highly reliant identification of fundamental interrelations within materials for
characterization and optimization in a scale-bridging manner [15].
For more detailed information on recent ML applications in materials science we refer to the general reviews of
Mueller et al [109], Wagner et al [162], Dimiduk et al [33], or Wei et al [168]. Examples for successful applications of
ML techniques in materials science are, for example, to represent inorganic materials [61, 134, 135], predict fundamen-
tal properties [29, 79, 100], create atomic potentials [80], identify functional candidates [81, 89, 94], analyze complex
reaction networks [157], or guide experimental design [71, 122], high-throughput phase diagrams and crystal structure
determination [49].

2.1.1 Open problem—interpretability

However, one of the major criticisms of ML algorithms in science is the lack of novel understanding and knowledge arising
from their use. This is mostly because more complex ML algorithms are often treated as black boxes. Those machine-built
models are hard to understand for humans [133].
For a better acceptance of ML models, data scientists aim to establish clear causal relations between materials structure
defined broadly across length scales and properties. Especially scientific models have further constraints such as a mini-
mal number of parameters and adherence to physical laws. It is the obligation of the data scientist to translate the results
of their work into knowledge other scientists can use in aiding for example materials discovery or deployment [162]. Use-
ful techniques for finding simple, reduced and interpretable models are for example principal component analysis (PCA)
[172], cross-validation and regularization, and a thoughtful choice of model.
PCA is a powerful technique for data dimensionality reduction. Large datasets are increasingly widespread. In order
to interpret such datasets, PCA can be applied to drastically reduce their dimensionality in an interpretable way, such that
most of the information in the data is preserved [67]. PCA extracts the orthogonal directions with the greatest variance
from a dataset, the resulting principal components being linear combinations of the original variables. However, principal
components are not necessarily simple to interpret physically but as the extracted features are linear combinations of
the original variables they can still be intuitively explained. Moreover, it allows a very straightforward data visualization
through data projection onto the main extracted components [158]. However, PCA might be the wrong choice if features
are not covariant.
Another way to achieve interpretable ML models is intelligent feature selection for dimension reduction and thus
easier interpretability. Regularization of a model entails adding a tunable penalty on model parameter size to the cost
function being minimized leading to a reduced feature space [162].
Furthermore, the choice of ML model has an immediate impact on its explainability. Regressions lead to coef-
ficients whose size gives information about the relative size effect of modifying an input on the output. Decision
trees are set up like flow charts and therefore easy to read. More complex models such as artificial neural networks
(ANNs) are missing a clear explanation of the machine’s “thinking” due to complex node interactions. But meth-
ods such as feature visualization [37, 142] or attribution [44, 150, 176] exist, which allow a better understanding and
interpretability of black box models. However, sometimes it might be reasonable to trade model accuracy for better
explainability.

2.1.2 Open problem—small data

Contrary to the curse of dimensionality, there is often the problem of small data in material characterization, because the
experimental or simulative generation of data is complex and expensive. Also the availability of the testing units limits
data generation. Special test environments and material properties that are difficult to implement in the laboratory are
often required, such as the creation of a corroded workpiece.
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
4 of 21 STOLL and BENNER

Often models are fitted to extremely small training sets which does not play to the strength of ML and will not allow
the replication of the success ML methods had in other fields. It is of course possible to use ML methods as a simple fitting
procedure for small low-dimensional datasets [133].
A few approaches are known to tackle this problem. For example, a ML model can be constructed by restricting the
configurational space of materials, such as predicting the band gaps of selected families of semiconductors with fixed
composition or crystalline structure instead of modeling compounds spanning a wide chemical space [32, 119, 179].
Another approach by Zhang et al [179] proposes to incorporate the crude estimation of the property in the feature space
to establish ML models using small sized materials data, which increases the accuracy of prediction without the cost of
higher degree of freedom.
Another approach for insufficient training data is the additional integration of prior knowledge into the training
process, which leads to the notion of informed ML [161], or more specifically physics-informed ML [124, 177, 180].
Domain knowledge is often given as a set of additional constraints [149]. The integration of additional domain knowl-
edge generates a new hybrid formulation of the ML problem which then ideally leads to physically meaningful and
significantly more accurate interpretations of the data [38]. Besides adding constraints, expert knowledge can be incor-
porated in different ways. Up to recently, it was mostly limited to labeling data for supervised learning and setting prior
probabilities in Bayesian networks [90]. However, in semi-supervised clustering applications, user guidance can be given
by partial labeling information, which can be incorporated using hard constraints [24]. The domain knowledge can also
be integrated into the process of building a ML model by the application of data visualization which often improves the
accuracy of the resulting model [90]. Another approach is the monotonization of ML functions based on known physical
relationships [31].
In general, ML systems are rarely viewed in the context of small data, where an insufficient data size for the training
model compromises the learning success. The bottleneck of the database size especially limits applications, in which the
construction of a database via experiments is time consuming and costly [136]. Thus, the recent development of materials
databases might be helpful in tackling the small data problem.

2.1.3 Existing databases

With the launch of the materials genome initiative [155] in 2011 and the coming of the big data era, a large effort has been
made in the materials science community to collect extensive datasets of materials properties and to provide materials
engineers with ready access to the properties of known materials. Existing databases are the materials project [63, 98],
the inorganic crystal structure database [39, 126], the Materials genome initiative [155, 171] ,the NOMAD archive [34,
113], the Topological Materials Database [156], Supercon [97], or the National Institute of Materials Science 2011 [111]
with many databases of material properties of metal alloys, or the National Institute of Standards and Technology [112]
with databases of properties for material classes such as structural ceramics, oxide glasses, superconductors [50, 73, 121,
127, 167, 175]. A more comprehensive list of material databases can be found in Correa-Baena et al [27].
Traditionally, negative results are often discarded and left unpublished. However, negative data are often just as impor-
tant for ML algorithms as positive results in order to prevent bias in the data. In some disciplines with a longer tradition
of data-based research (like chemistry), such databases already exist. In a similar vein, data that emerge as a side product
but are not essential for a publication are often left unpublished producing so-called publication bias [108]. This eventu-
ally results in a waste of resources because other researchers then have to repeat the work in order to produce a balanced
dataset for ML applications [86].
However, few of these available databases are ready to use with informatics techniques because they lack the uniform
data formats or application programming interfaces required for informatics software [167].

2.1.4 Materials informatics

The databases mentioned above contain information on numerous properties of known materials and are essential for
the success of materials informatics [86]. For more general information, the reader is referred to [145], which gives an
introducing review on materials informatics, that aptly describes the concept of big data in materials science. Another
publication on materials informatics is, for example, [169] which introduces four main research areas in materials infor-
matics: standardization of representation and exchange of material data; organization, management, retrieval, filtration
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
STOLL and BENNER 5 of 21

and correlation of material data; material graphics; and data mining and knowledge discovery of material data. Another
review on materials informatics mostly focuses on atomic-scale modeling [167]. But it also promotes the idea to expand
materials databases to make more data easily accessible to informatics. The widespread use of such data requires the
digitalization and structuring of materials data. The data must also be easily sharable and accessible. Services that provide
software interfaces to allow for automated data querying, processing, and access are evolving, for example, the Materials
Data Facility [14] and Citrination [47, 116]. Wagner et al [162] propose a workflow for a materials informatics problem
focusing on (a) the assembly of primary features, (b) the construction of an exploratory model, (c) refinement of the model
to satisfactory accuracy, and (d) final training and deployment. By proceeding in an iterative fashion upwards in complex-
ity, the final model will be as simple as possible improving its explainability and interpretability. The shorter, but more
philosophical report of Rajan et al [125] focuses on the role of materials informatics that allows one to survey complex,
multiscale information in a high-throughput, statistically robust, and yet physically meaningful manner.

2.2 ML for metallic material characterization

Mechanical material properties are characteristics to be precisely predicted and controlled as they are strongly linked to
and highly affected by process parameters and resulting microstructures [15]. The basic idea of using ML methods for
material property prediction is to analyze and map the relationships (nonlinear in most cases) between the properties of
a material and their related characteristics by extracting knowledge from existing experimental or simulated data [86].
Mechanical behavior in simulations is often described by means of constitutive equations [15].
Research on the macroscopic performance of materials mainly focuses on the structure-activity relationship between
the macroscopic (eg, mechanical and physical) properties of a material and its microstructure [86]. Many material parame-
ters can be estimated to within an order of magnitude using elementary physical ideas. Whenever these parameters cannot
be reliably estimated as such, ML approaches can be helpful which then require experimental or simulative data [88].
Experimental testing methods which can be used on metals help understand materials and their properties better.
Typical destructive tests are bend test, impact test, hardness test, tensile test, fatigue test, corrosion resistance test, or wear
test, see for example [104]. In the following efforts of the materials community to enhance such tests and their results
with ML methods will be discussed.

2.2.1 Corrosion

Corrosion detection and monitoring are essential diagnostic and prognostic means for preserving material “health” and
reducing life-cycle cost of industrial infrastructures, ships, aircrafts, ground vehicles, pipelines, and so on [6]. More
recently, ML approaches have shown great potential to improve corrosion detection [66]. They will aid a human inspector
and significantly cut down on the time and cost associated with inspecting for example civil infrastructure and eliminate
the need for dependence on prior knowledge and human effort in designing features [12]. Popular approaches are the
application of ANNs for image processing-based corrosion detection.
Convolutional neural networks (CNNs) were used for corrosion detection by investigation of images and identifica-
tion of rusty parts in the image [12]. The proposed CNN outperforms state of the art vision-based corrosion detection
approaches that are developed based on texture and color analysis using a simple multilayered perceptron network.
Model input is an image of the material and region of interest, output is the classification information: corroded/not
corroded for a sliding window over the image. Overfitting due to small datasets can be avoided by using pretrained net-
works. Another example for the application of CNNs for corrosion detection is proposed by Bastian et al [13]. Similar
to Atha et al [12] images are used as input and classified into any of the four classes: no corrosion, low-level corrosion,
medium-level corrosion, and high-level corrosion. More investigations of similar CNN-based publications for corrosion
detection are listed in [13].
Fang et al [41] proposed a novel hybrid methodology combining genetic algorithms and support vector regression,
which is capable of forecasting the atmospheric corrosion depth of metallic materials such as zinc and steel. This hybrid
approach is capable of solving nonlinear regression estimation problems in materials science. The genetic algorithms
are adopted to automatically determine the optimal hyper-parameters for support vector regression. The inputs for the
support vector regression are temperature, time of wetness, exposure time, sulfur dioxide concentration, and chloride
concentration, respectively. The outputs are the predicted corrosion depth for zinc or steel.
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
6 of 21 STOLL and BENNER

Another support vector machine (SVM) approach was applied by Hoang et al [54] for image processing-based detec-
tion of pipe corrosion. The image texture including statistical measurements of image colors, gray-level co-occurrence
matrix, and gray-level run length is employed to extract features of the pipe surface. SVM optimized by differential flower
pollination is then used to construct a decision boundary that can recognize corroded and intact pipe surfaces by block-
wise classification of the original image. Also with the application of a support vector regression and based on a much
smaller database (trained on only 46 samples), the corrosion rate of 3C steel in different environments was predicted
based on five different seawater environment factors, including temperature, dissolved oxygen, salinity, pH-value, and
oxidation-reduction potential. The prediction error was very small [170].
Jimenez et al [66] compare various ML approaches (ANN, SVM, classification tree, and k-nearest neighbor) for
automatic pitting corrosion detection in 316L stainless steel. Model input are environmental variables such as chloride
concentration, pH, and temperature while the output is an information about the material being corroded or not. The
models based on ANNs and SVM with linear kernel were demonstrated to be a valuable tool to be applied for this purpose.
The classification performance for ANN and SVM is much better compared to k-nearest neighbor and classification tree
models for this application. The principal advantage compared with the traditional techniques is that it is not necessary
to apply a surface analysis technique to study corrosion behavior of a material.

2.2.2 Fatigue

Fatigue as the weakening of a material caused by cyclic loading that results in progressive and localized structural dam-
age and the growth of cracks can also be predicted with the incorporation of ML methods. The prediction of fatigue in
welded structures for a wide range of structural materials by multiscale FEM and ML was proposed by Shiraiwa et al [140].
Two ML algorithms are applied: one is deterministic ML based on the traditional methods, and the other is model-based
ML. Deterministic ML such as multivariate linear regression and ANNs use chemical composition, processing parame-
ters (reduction ratio, heat treatment), inclusion sizes, and fatigue strength as input features to accurately predict fatigue
strength. For model-based ML, microstructures and stress-strain curves in 40 low carbon steels with different chemical
compositions and heat treatment conditions were prepared to create the learning dataset which was used to train an ANN.
This approach allows for incorporation of prior knowledge of structure and property, and it can account for uncertainty
such as scattering of fatigue life.
Another attempt to identify novel connections between fatigue properties and a variety of material parameters with the
help of ML algorithms has been made by Agrawal et al [8] for the prediction of fatigue strength of steel from composition
and processing parameters such as chemical composition, upstream processing details, heat treatment conditions and
mechanical properties. Various ML methods such as basic regression, decision trees, SVM, ANN were used. Most success
showed ensemble methods and individualized methods for different types of materials.
Machine learning techniques have also been utilized to predict material fatigue life for P91 steel base metal based on
the hold time in fatigue tests [178]. A combined approach of genetic algorithms and SVM is used to predict the fatigue
life with high accuracy.
Abdalla et al [2] use an ANN radial basis function model, taking the maximum tensile strain and pressure ratio as
input, and put forward the model of fatigue life of steel reinforcing bars.

2.2.3 Creep
Creep is a type of metal deformation that occurs at stresses below the yield strength of a metal, generally at elevated
temperatures. Creep rupture is becoming increasingly one of the most important problems affecting behavior and per-
formance of power production systems operating in high temperature environments and potentially under irradiation as
is the case of nuclear reactors. Creep rupture forecasting and estimation of the useful life are required to avoid unantici-
pated component failure and cost ineffective operation [25]. The material behavior is influenced by the multidimensional
interdependencies between the individual elements of the chemical composition, the heat treatment parameters, product
form, tensile properties and microstructure, which are difficult to describe using simple analytical methods. Modeling
with ML techniques therefore seems to be an interesting alternative. Moreover, the application of ML takes away the
requirement for long and expensive experiments [46].
For the design of materials, creep is considered an important material property. However, quite often such designs
only focus on one objective (eg, creep) without considering the comprehensive design of multi-property [164]. For the
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
STOLL and BENNER 7 of 21

investigation of creep rupture life and rupture strength of austenitic stainless steels [146] once again ANNs are popular
models. For the prediction of creep rupture life and the creep rupture stress for a given stress, the training database
for the input parameters contains test conditions (stress and temperature), chemical composition, solution treatment
temperature, and time (the latter being available in a very limited number of cases), nature of the quench following, grain
size, and logarithm of ruptured life for a broad variety of stainless steels.
Chatzidakis et al [25] employ and compare general regression neural networks, ANNs and Gaussian processes to
capture the underlying trends and provide creep rupture forecasting. Input parameters are experimental creep rupture
data. However, the overall performance of the developed models was insufficient.
In the study of Shin et al [139], the five different ML models random forest (RF), linear regression, k-nearest neighbor,
kernel ridge, Bayesian ridge are applied for the prediction of Larson-Miller parameters which represent the creep behavior.
From a wide range of available features (466), relevant ones are selected with optimization approaches and different
set-ups of features and models evaluated. Highest accuracy was obtained by RF for a varying number of top ranking
features between 5 and 21.
For the prediction of rupture and creep rupture stress of 9%Cr steels, a multilayer perceptron neural network
was applied with the input parameters chemical composition, heat treatment information, geometrical form of the
investigated components [46].

2.2.4 Flow behavior and work hardening

In sheet metal forming operations, mechanical properties of the sheet material such as flow stress or stress-strain curves
greatly influence metal flow and product quality [52]. The flow stress can be determined by tensile tests which provide
uniaxial information about stress-strain behavior. Hardness tests measure the resistance of material to an indenter, and
hardness correlates well with flow stress. Compression testing reveals flow stress but is complicated by friction and buck-
ling. Bend testing and torsion testing provide a good measure of flow stress and fracture resistance but are complicated
by radial stress and strain gradients [173].
One of the main advantages of applying ML approaches is that it is not necessary to postulate a mathematical model
at first, which is quite difficult because of the nonlinearities in the response of the deformation behaviors of the materials
under elevated temperatures and strain rates and the factors affecting the flow stress. Lin et al [82] propose a feed forward
back propagation ANN model to predict the constitutive flow behaviors of 42CrMo steel during hot deformation, and
investigate the general nature of the influence of strain, strain rate and temperature on the compressive deformation
characteristics of 42CrMo steel. The capability of the developed ANN model to predict the flow stress level, the strain
hardening, and flow softening stages is also investigated. The inputs of the ANN are deformation temperature, log strain
rate, and strain whereas flow stress level, the strain hardening, and flow softening stages are the output. Low absolute
relative errors for training and test data promise a good generalizing model.
Gupta et al [51] predict flow stress based on a feed forward ANN trained with the back propagation algorithm. Input
features were strain, strain rate, and temperature. The implemented ANN produced more accurate results than conven-
tional mathematical models such as Johnson Cook, modified Zerilli-Armstrong, and modified Arrhenius. support vector
regression based flow stress prediction for austenitic stainless steel 304 with strain, strain rate and, temperature as inputs
and the flow stress as output was proposed by Desu et al [30] and provides more accurate results than the conventional
mathematical models.
The identification of work hardening properties of steel and an aluminum alloy from indentation tests is described
by Meng et al [101]. The authors propose a material parameter identification protocol based only on the imprint shape
of the indentation test using orthogonal decomposition and manifold learning for the prediction of the strain hardening
exponent and yield stress.
ANN modeling for anisotropic mechanical properties and work hardening behavior of Inconel 718 alloy at ele-
vated temperatures was proposed by Mahalle et al [91]. Strain and temperature are used to successfully predict material
properties such as ultimate strength, yield strength, elongation, and strain hardening coefficient.

2.2.5 Tensile properties

Tensile properties indicate how the material will react to forces being applied in tension. Determining the tensile prop-
erties is crucial because it provides information about the modulus of elasticity, elastic limit, elongation, proportional
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
8 of 21 STOLL and BENNER

limit, reduction in area, tensile strength, yield point, yield strength, and other tensile properties [123] which then define
the state of the material, its longevity, or its ability to perform in an application. Thus, the accurate prediction of tensile
properties has great importance for the service life assessment of structural materials [163].
The ultimate tensile strength (UTS) of iron castings which gives information about the capacity of a metal to resist
deformation when subject to a certain load was predicted based on a variety of input features such as composition, size
of casting, cooling speed or thermal treatment (25 variables) to gain information about the mechanical properties of a
foundry and thus predict foundry defects [131]. Estimating the value of UTS is one of the hardest issues in foundry pro-
duction, due to many different circumstances and variables that are involved in the casting process. Bayesian networks,
k-nearest neighbor, and ANNs were used for classification of the UTS. All of the investigated approaches perform well,
but ANNs outperformed the other classifiers. Similar to the aforementioned study, Sterjovski et al [148] also used ANNs
(of back propagation type ANN) to predict mechanical properties of steel, such as the impact toughness of quenched
and tempered pressure vessel steel exposed to multiple postweld heat treatment cycles; the hardness of the simulated
heat affected zone in pipeline and tap fitting steels after in-service welding; and the hot ductility and UTS of various
microalloyed steels over the temperature range for strand or slab straightening in the continuous casting process. Input
parameters were composition, cooling rate, temperature, and thickness. It was shown that ANNs could successfully pre-
dict all mechanical properties investigated. Another example for the application of ANNs is proposed by Sankar et al [130]
who predict elongation, self-tempering temperature, and yield strength for reinforcement steel bars subjected to thermo-
mechanical treatment based on two input parameters (bar diameter and quenching duration). The numerical results of
the ANN are compared with experimental results and are found to be in good agreement.
Pruning and predator prey algorithms were applied by Datta et al [28] which were able to extract more knowledge from
the input data, than typically possible with conventional ANN analysis. Alloy composition and the thermo-mechanical
controlled processing parameters, deformation in different temperature zones, finish rolling temperature, and cooling rate
of high strength steels have been taken as input parameters, whereas UTS, yield strength, and percentage elongation were
predicted. It was shown that in this type of steel the yield strength depends mostly on the solid solution hardening and
the microstructural constituents while UTS is more influenced by the precipitation hardening, but all these strengthening
mechanisms have a negative effect on the ductility of the steel.
Pattanayak et al [117] investigated the role of the composition and processing parameters on the mechanical prop-
erties of API grade microalloyed pipeline steel, in respect to its strength, impact toughness, and ductility. ANN models,
capable of prediction and diagnosis in nonlinear and complex systems, are used to obtain the relationship of composi-
tion and processing parameters with said mechanical properties. Then the models are used as objective functions for
the multiobjective genetic algorithms for evolving the tradeoffs between the conflicting objectives of achieving improved
strength, ductility and impact toughness. The Pareto optimal solutions were analyzed successfully to study the role of
various parameters for designing pipeline steel with such improved performance.
For the ANN-based prediction of yield strength, UTS, ductility (elongation, and reduction of area) of ferritic steel
weld metals appropriate for the welding of high strength low alloy steels, Metzbower et al [103] use the chemical com-
position of as deposited weld beads, and the cooling rate. The various established ANN models are found to work well
once combined to an ensemble of best models. The ensemble produced a more reliable prediction than an individ-
ual model and reproduced known metallurgical trends well. Poudel et al [120] also compare different ML models and
ensembles for the prediction of UTS, yield strength, elongation of steel bars instead of performing tensile tests. Input
parameters were process parameters from the rebar manufacturing process such as for example material composition,
temperatures, rod diameter, rod speed, and cooling rate. Investigated classification algorithms were among others multi-
variate linear regression, principal components regression, partial least squares regression, ANNs, and locally weighted
regression. The main conclusion from this research is that any single model cannot efficiently describe the complicated
relationship of the input-output space of the rebar manufacturing process. Ensemble methods, however, as combination
of ML approaches used together in conjunction with some model selection or model weighting techniques lead to robust
prediction systems.
Applications of traditional ML algorithms for the prediction of tensile properties were proposed by [137, 138, 151, 163,
164]. Shigemori et al [137, 138] reported about the successful application of locally weighted regression in predicting the
tensile strength for a certain type of steel product which is produced by hot rolling. As input, 18 items were selected from
chemical composition, heating, rolling, and cooling temperature. These variables have a clear physical causal relationship
with the output variable. Furthermore, least squares SVMs are suitable approaches for the prediction of the elastic mod-
ulus and yield stress of materials. In [151] FEM-simulated load-indentation curves of Al6061 and Al7075 are investigated
for the determination of these material parameters based on a training set of large strain-large deformation FEM for the
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
STOLL and BENNER 9 of 21

simulation of indentation tests. Characteristic features are extracted from the load-indentation curves and used in the ML
model. The proposed least squares SVM model is capable of predicting reasonably accurately the elastic modulus and yield
stress of materials based on the load-indentation curves of dual conical indenters with different half-angles. For material
design, a RF model in combination with an optimization algorithm was found to relate yield strength, impact toughness
and total elongation with material composition information and treatment parameters for the production of RAFM steels
[163, 164]. For yield strength, highly correlated features were tempering temperature and C content; and tempering time
and Cr content for elongation. The accuracy and generalization ability of the RF were acceptable (R2 > 85%).
An interesting study different from most approaches using material composition information as input features for
material parameter prediction was proposed by Fragassa et al [45] for the prediction of the tensile behavior of cast alloys
such as yield strength, ultimate strength, ultimate strain, and Young’s modulus, by a pattern recognition analysis on exper-
imental data and the application of RF, ANN, and k-nearest neighbor. All information is directly taken from micrographs.
For the prediction of UTS and yield strength ANNs show the best results.
Another application of ANNs is the interpretation of acoustic emission data for failure prediction. Christopher et al
[26] propose the prediction of the ultimate strength of aluminum/silicon carbide (Al/SiC) composites by using acous-
tic emission parameters through ANN analysis. This approach was earlier pursued for the prediction of the ultimate
strength of unidirectional T-300/914 tensile specimens using acoustic emission response and an ANN back propagation
algorithm [132].
To prevent surface cracks on cast steel, its hot ductility must be monitored. Experimental investigations are diffi-
cult to execute. Thus, ML models are proposed for the prediction of hot ductility. For example, a multivariate linear
regression [85] predicted hot ductility and grouped 12 chemical elements that had similar experimental effects on duc-
tility. The cooling condition used here is different from actual continuous-casting conditions, so the model is difficult
to apply in commercial setups. A back-propagation ANN method has been used [148] to predict hot ductility for vari-
ous microalloyed steels over the temperature range for strand or slab straightening in the continuous casting process.
However, the recorded data are limited and do not lead to a generalizing model. Additionally, the ANN used only one
hidden layer, so it has a limited ability to describe the complex relationship between input and output. Also an ANN
model was used to predict high-temperature ductility of various steel grades from their composition and thermal his-
tory (described by five experimental variables) [78]. The developed model can predict ductility for a wider composition
range and thermal history than previous studies have achieved. Therefore, it can be used effectively in commercial
production.

2.3 ML for the SPT

As mentioned above, the estimation of tensile properties of materials such as elastic modulus, yield strength, or strain
hardening exponent is considered to be of fundamental importance. Conventional tests are destructive in nature and
require reasonable specimen cross-section and volume. There are situations where limited volume of material is avail-
able for property assessment, like in material development, failure analysis and remaining life assessment of in-service
components, materials for pressure vessels, turbines, thermal power plants, or chemical processing industries [9, 60]. In
these situations, test techniques using small volume specimens become more attractive. Small specimen test techniques
have been established as a reliable alternative to the traditional tensile test, as the results of these test techniques are in
good agreement with the tensile test and are reproducible when tested under controlled conditions [11].
The extraction of mechanical properties of in-service materials through such small specimen test techniques has
become more popular in recent years. Of the available small specimen test techniques, SPT methods have proven to be
promising [11, 92, 93, 165]. In the SPT, a thin disk like specimen of small diameter is deformed in a miniaturized deep
drawing experiment. The measurable output is the load displacement curve of the punch, which contains information
about the elasto-plastic deformation behavior and about the strength properties of the material [3]. Such test methods
are basically nondestructive in nature and are proficient enough to extract the flow properties of the materials only using
small volumes of material. After more than 40 years’ development, there are some standards in SPTs, such as ASTM-F2183
[147], GB/T-29459 [143], and CEN CWA-15627 [22].
A prerequisite for using this test is to establish correlations between SPT and conventional tests [10, 65], such as
relations between the tensile test and the Erichsen cupping test for the determination of the critical damage value curves,
the initiation time and location of fracture [72]. Other examples are the correlation between the SPT and tensile tests for
the estimation of the tensile strength [74] or the construction of stress-strain curves only from indentation tests [174].
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
10 of 21 STOLL and BENNER

Here the link between mean pressure and the indent diameter, obtained from indentation tests, to the stress and strain
determined from a tensile test was obtained by the empirical relation of Tabor [152].
Exploiting force-displacement curves (FDCs) from spherical indentation simplifies the identification of mechanical
properties. Parameters such as hardness, deformation mode, yield stress, Young’s modulus of the indented material can
also be extracted out of the imprint shape, which is a valid alternative for using the indentation curve [102]. For example,
Milivcka et al found linear correlations between the SPT maximum force and the tensile strength of a 9% Chromium steel
for creep resistant high temperature applications [105].
Detailed analyses of stress and strain in the SPT disc have been performed by means of analytical elastic-plastic model-
ing, see for example [20, 23] and by FEM to underpin the empirical correlations [4, 21, 141]. FEM simulations are mainly
used to generate FDCs based on an assumed material constitutive law which are then compared with experimental FDCs.
This approach is called inverse analysis [83, 87]. For more information, see for example [118, 168]. However, the solution
to the inverse identification problem is nonunique. Furthermore, there are other problems such as insufficient accuracy,
indentation frame/machine compliance, noisy input data, difficulty in determining the exact starting point of the load
(force)-displacement curve, bending of the specimen in thin-sheet indentation and FE mesh dependence [102].
Various mechanical properties that can be extracted through SPT are for example fracture mode [69, 99, 106], yield
stress [21, 70, 84], UTS [59, 96, 105, 165] (based on hydraulic bulge tests which are similar to SPT but high-pressure
hydraulic oil is used instead of punch to cause specimen deformation); Young’s modulus [42, 57, 114, 153], fracture
toughness [4,18,76,83,95,107,166], creep properties [36, 48], or elastic plastic properties [43, 128].
These examples show that the SPT is a widely used approach for the determination of various material parameters
from a small amount of material. However, the approach also has a few disadvantages, which have to be kept in mind.
To begin with, the selected sample size may not represent the bulk material. The sample size effect has an important
influence on the mechanism of the fracture of SPT, because often the thickness of the SPT sample is only 5 to 8 times of
the mean grain size. The changes in grain-boundary distribution and the grain orientation, that resulted from the sample
size effect, may affect the mechanical behaviors of the SPT samples. These parameters may affect the mechanism of
fracture on small scale material samples [144]. In addition, the SPT response is sensitive to various test parameters such
as specimen shape, specimen thickness, test speed, ball diameter, clamping force, and material. For more information,
the reader is referred to [10], where an extensive collection of SPT configurations is listed.
In recent years, using ML approaches on SPT data has become popular. Most approaches incorporate ANNs for the
identification of material properties from SPT data. Abendroth et al [5] identify ductile damage and fracture parameters
from the SPT using ANNs. FEM is used to compute the load displacement curves. Via a systematic variation of the material
parameters a data base is built up, which is used to train the ANNs. This neural network can be used to predict the load
displacement curve of the SPT for a given material parameter set. The identified material parameters are validated by
independent tests on notched tensile specimens. A similar approach was followed by Abendroth et al [3, 4] earlier for
the estimation of the hardening and damage parameters of high strength steels. The combination of FEM, SPT data, and
ANNs was used to identify the parameters of the Gurson-Tvergaard-Needleman model for ductile damage and fracture
parameter prediction.
Linse et al [83] use synthetic load displacement curves generated via FEM for a variety of material parameters. This
database is then used to train ANNs, which approximate the load displacement curves of the SPT as a function of the
material parameters. The identification procedure itself consists of an optimization algorithm, minimizing the differ-
ence between the measured load displacement curves and its approximation by the neural networks until the true set
of material parameters is found. Prerequisite is an accurately working FEM-Model. The approach was applied for the
identification of hardening parameters and Weibull-parameters in the brittle and brittle-ductile transition region of two
nonirradiated reactor vessel steels.
Another employment of FEM simulations of tensile, bulge, Erichsen tests is proposed by Abbassi et al [1] for the
calculation of damage parameters. The Gurson-Tvergaard-Needleman model was employed. An identification procedure
based on ANNs is used to determine the material parameters of the Gurson-Tvergaard-Needleman model. The ANN was
trained by using the FEM results of the notched tensile test with varying the damage parameters. A good efficiency for
the identification of damage parameters was proven. The Poisson ratio can also be predicted using a similar approach
(combination of SPT and FEM and ANN) [56]. Using FE simulations, the relation between the material parameters and
the quantities characterizing the depth-load response is calculated. An approximate inverse function represented by an
ANN is derived on the basis of these data.
In contrast to all the ANN-based approaches described so far, the study of Meng et al [102] suggests the use of man-
ifold learning for indentation-based material characterization. However, FEM was also used there. Input parameters
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
STOLL and BENNER 11 of 21

were features describing the imprint shape of SPTs. The manifold learning approach was able to iteratively reduce the
distance between the FE-simulated and the experimental imprint shapes in order to identify the material hardening
parameters. The approach was successfully shown for three different materials: AISI1095 steel and two aluminum alloys
EN AW-2017F and EN AW-5754F.
Another innovative approach suggests the application of digital video processing of the forming process created by
the Erichsen cup test [129]. The OpenCV library [16] was used to develop a sheet-metal-forming image analysis program
which identifies the indent shape and position to calculate the anisotropic coefficient which showed good agreement with
tensile test results. However, results were biased by the environment around the testing area.
A code of practice on SPT was established in 2010 in order to harmonize the various test set-ups and to achieve a better
comparability of the results of different labs [58].

3 AN APPLICAT ION O F ML IN MATERIAL PRO PERTY PREDICTION

Similar to the studies mentioned above, we want to apply ML models to extract material properties from SPT data since
the SPT has long been recognized as a supportive means for the development and monitoring of structural materials.
Conventionally, the following empirical correlation has been used for the estimation of the UTS 𝜎UTS from FDCs [17,
21, 35, 77, 96]

𝜎UTS = 𝛽UTS ⋅ Fm ∕(h0 ⋅ vm ) (1)

with F m being the maximum force, vm the corresponding punch displacement, h0 the initial specimen thickness, and
𝛽UTS an empirical coefficient. However, various researchers [9, 77] propose the determination of 𝜎UTS based on the force
F i instead of F m for the correlation with the UTS in order to avoid a strong dependence of the correlation factor 𝛽UTS
on the tensile properties of the material. F i = F(vi ) is the force at the intersection point vi of FDCs for varying uniform
elongations and constant UTS 𝜎UTS and constant initial flow stress 𝜎y0 .

𝜎UTS = 𝛽UTS ⋅ F(vi )∕(h0 2 ) (2)

with 𝛽UTS being geometry and material dependent. More information about this empirical approach can be found in [9].
In addition to the empirical approach (1), (2), we propose two ML models to determine the UTS 𝜎UTS from SPT data,
shown in Figure 1B.

3.1 Experiments

SPT data of three ferritic-martensitic (f/m) Cr-steels and two bainitic reactor pressure vessel steels [9] were used to predict
the UTS 𝜎UTS based on ML approaches. The UTS 𝜎UTS was obtained by tensile tests. A materials overview can be found in
Table 1, which was adopted from Altstadt et al [9]. The P92 was available in the, as received condition, and in four different
heat treatments. The specimens were of 0.5 mm thickness. SPT was conducted with a punch diameter of 2.5 mm, a receiv-
ing hole diameter of 4 mm, edge size 0.2 mm, and the edge being beveled. All tests were performed at a displacement rate of
0.5 mm/min. The displacement v was measured by an inductive sensor with an accuracy of ±1 μm. The FDCs F(v) of SPTs
at room temperature are shown in Figure 2A (mean of all available curves per material at room temperature). The F(v)
curves of the material P91 for different temperatures are shown in Figure 2B. The curve for T = −177◦ C represents a brittle
failure, the other test curves represent ductile failure. By using nine different materials and heats, and by testing selected
materials at different temperatures, a wide variation of tensile properties is obtained. Table 2 gives the number of con-
ducted tests and their test temperatures for each type of material. Figure 1 shows the UTS 𝜎UTS for all investigated materials
and temperatures. For more detailed information on the experimental set-up the reader is referred to Altstadt et al [9].

3.2 Preprocessing

First, unified sampling points for the displacement were defined with a distance of 44 μm based on a sampling rate of
roughly 1 μm in the raw data. Then each of the displacement values is defined as a feature, leading to 151 features for
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
12 of 21 STOLL and BENNER

T A B L E 1 Materials overview
Material Type Product form Heat treatment Reference

P91 f/m steel Hot rolled pipe ø 360 × 50 mm Normalization 1040-1100◦ C/30 min [75]
tempering 730-780◦ C/60 min
P92 f/m steel Hot rolled pipe ø 219 × 22 mm AR: standard normalization + [55]
tempering
HT1: AR + 800◦ C/2 h
HT2: AR + 760◦ C/2 h
HT3: AR + 750◦ C/2 h
HT4: AR + 740◦ C/2 h
Eurofer97 f/m steel Hot rolled plate 14.5 mm Normalization 980◦ C/27 min/air [53, 154]
cooling
tempering 760◦ C/90 min/air cooling
22NiMoCr 37 Bainitic steel Reactor pressure vessel Biblis C 890◦ C/4 h/water quenching [181]
650◦ C/7 h/air cooling
15Kh2MFA Bainitic steel Reactor pressure vessel Original RPV manufacturing [160]
Greifswald technology
unit 8

T A B L E 2 Overview of SPTs
Material Number of tests Test temperatures

P91 23 −177 … + 331

P92-AR 3 Room temperature
P92-HT1 5 Room temperature
P92-HT2 4 Room temperature
P92-HT3 3 Room temperature
P92-HT4 5 Room temperature
Eurofer97 8 −24 … + 250
22NiMoCr37 31 −151 … + 332
15Kh2MFA 33 −150 … + 332

(A) (B)

FIGURE 1 A, UTS 𝜎UTS for the investigated materials and temperatures. B, Scheme
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
STOLL and BENNER 13 of 21

(A) (B)

FIGURE 2 A, SPT data at room temperature for materials listed in Table 1. B, SPT data for P91 at different temperatures

(A) (B)

FIGURE 3 A, Pairs plot of selected features, temperature and 𝜎UTS . B, Pairs plot of first five principal components, temperature, and 𝜎UTS

displacement values up to 1.5 mm. Failure occurred for displacements larger than 1.5 mm. An additional feature is the
test temperature. Now, each time series is considered to be an observation. Dimensions correspond to the number of
displacement sampling points + 1 (temp).
As the number of dimensions (p) increases, the volume of the domain increases exponentially. This, in turn, requires
more samples (n) from the domain to provide effective coverage of the domain for a learning algorithm. This problem was
introduced earlier as the curse of dimensionality [64]. ML algorithms overcome the curse of dimensionality by making
assumptions about the data and structure of the mapping function from inputs to outputs which adds a bias and leads to
a loss of the generalization power of the ML model [64]. Ways to approach this problem are feature selection, projection
methods, or the application of regularized algorithms.
Another problem when using FDC-data as input data for a ML algorithm is the multicollinearity of the features
extracted from these curves. Exemplary correlations between features 1, 5, 45, 95, 150 and the temperature are shown
in Figure 3A. Subsequently, PCA is performed on the FDC-data in order to eliminate colinearity in the features. The
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
14 of 21 STOLL and BENNER

T A B L E 3 RMSE comparison
Empirical LM RF

RMSE 44.1 41.2 40.4

(A) (B) (C)

FIGURE 4 A Empirical 𝜎UTS vs experimental 𝜎UTS . B, LM based 𝜎UTS vs experimental 𝜎UTS . C, RF based 𝜎UTS vs experimental 𝜎UTS

broad idea behind this scheme is that, in contrast to the original features, the principal components will be uncorrelated.
Furthermore, one expects that a small number of principal components will explain most of the variance and therefore
provide an accurate representation of the dataset [133]. For the present dataset a total of six principal components explain
99% of the overall variability. Now, the pairs plot of the new features in Figure 3B shows only horizontal or vertical lines
between input features (principal components) which is typical for uncorrelated features.

3.3 ML models

For the estimation of tensile properties such as the UTS 𝜎UTS from SPT data, empirical equations such as (1) or (2) can be
used. The force F i can be associated with the onset of plastic instability and is therefore well suited for a correlation with
the UTS 𝜎UTS of the uniaxial tensile test as proposed in [9]. With the number of available data points being rather small
a linear regression model (LM) in combination with PCA and a RF based on the original SPT data are proposed for the
prediction of 𝜎UTS . The root mean squared error was used to compare the different approaches.
For the evaluation of the performance of the ML approaches 10-fold cross validation was conducted with the results
shown in Table 3. Figure 4A-C shows the accuracy of the empirical approach, LM and RF, respectively.

3.4 Application conclusions

The application presented here aims at the determination of the tensile property 𝜎UTS from SPT data. Altstadt et al [9]
proposed an empirical equation which correlates the punch force at onset of plastic instability of the SPT with 𝜎UTS .
The correlation factor 𝛽 for the estimation of 𝜎UTS needs to be determined before the empirical equation can be applied
and 𝛽 is dependent on the sample geometry and the investigated material. The application of ML models on SPT data
works well for the prediction of 𝜎UTS , as Table 3 shows. However, the amount of available data is very limited and the
generalization power of the LM and RF for other materials, temperatures, geometries has to be critically evaluated. RFs
usually do not generalize well. Thus, the application of RF on SPT data might not be sensible for 𝜎UTS < 500 MPa or 𝜎UTS >
1000 MPa. More experimental data will have to be collected to verify the RF model further. However, the interpretation
of the RF based on SPT data is straightforward and leads to relevant information about the importance of features for the
performance of the model.
The LM approach is quite similar to the empirical approach of Altstadt et al. However, not only one specific force value
of the SPT is used to correlate to 𝜎UTS but the combination of several principal components generated from the SPT curves
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
STOLL and BENNER 15 of 21

is employed which might justify the slightly higher accuracy of the LM approach. Thus, LM is a great tool to analyze the
relationships among the variables but in combination with PCA the interpretability of the resulting model is not intuitive.
Overall, correlations between SPT curves and 𝜎UTS clearly exist and SPT data can be used for the prediction of tensile
properties such as 𝜎UTS for five different steels (nine heats). Once the model is established, there is no further need for
the conduction of time consuming tensile tests. In the future, the models might have to be re-established on a broader
database. For the determination of 𝜎UTS , FEM-simulations might not be necessary anymore. However, for more detailed
information about the underlying constitutive laws via inverse analysis, FEM simulations will likely remain important
an inevitable.

4 DI S C USSION AN D CON CLUSIONS

This survey addresses applications of machine learning strategies in materials science for material characterization. There
exists a wide range of promising applications for ML in materials science, for example, material discovery, molecular
dynamics, and global structural prediction. The demand for new approaches dealing with limited data is huge.
It was shown that data-driven approaches play a significant role in materials research in order to find relationships
between the structure of a material and its properties. These relationships are often not linear. It is difficult to find generic
patterns among multiple length scales and timescales. With experiments only, this cannot be achieved [125]. Therefore,
data-mining techniques are indispensable for the recognition of correlations in the (experimental and simulated) data.
As the amount of publicly available materials data grows, ML techniques in particular will be able to extract from these
data sets scientific principles and design rules that could not be determined through conventional analysis [62].
The majority of early ML applications to materials science employed straightforward and simple-to-use algorithms,
like linear kernel models and decision trees. Now, these proofs-of-concept exist for a variety of applications even though
there is a lack of benchmarking datasets and standards [133]. To date, ML often cannot realize the expected accuracy when
applied to some tasks due to insufficient material data. Therefore, a more accurate model that was trained on a small but
accurate data set is only meaningful within the input data space but does not generalize well while a less accurate model
on a wide input data space is better at generalization but less exact. Therefore, accelerating the construction of publicly
accessible material databases is highly important for the future development of ML in materials science [168]. Another
issue that holds back the development of precise ML models is the absence of failure data. In this case, a cultural shift
toward the publication of all valid data, may it be positive or negative, is required [133, 168].
The majority of ML approaches in materials science is based on ANNs. However, conventional ANNs still suffer from
several weaknesses such as the need for a large number of controlling parameters, the difficulty in obtaining stable solu-
tions, the danger of overfitting and thus the lack of generalization capability [41]. However, ANNs have been enormously
successful in understanding complex materials behavior, such as mechanical behavior (flow stress, hardness, tensile
strength, fracture strength, and fatigue behavior) of metal alloys subjected to certain heat treatment and/or deformation
procedures, as well as in the prediction of micro-structures and phases resulting from heat treatment and/or deformation
processes. The most practical way to capture the complex dependence of a desired macroscopic property on the various
process parameters is through such learning methods [109]. ANNs have the potential to minimize the need for expensive
experimental investigation and/or inspection of structural materials used in various applications, hence resulting in large
economic benefits for organizations [148].
In addition to ANN hybrid ML models or ensemble methods work well [41]. For this, multiple independent models
are built and the final regression or classification result is usually obtained as an average over the ensemble. In this way,
additional noise is introduced into the fitting process and overfitting is avoided [133]. However, there does not exist an
overall solution that can be considered the best. The most appropriate model has always to be found specifically for the
application and data situation.
Furthermore, this paper focuses on ML based material property prediction from SPT data. Such simple material tests
have gained popularity over the last couple of years because even though they are cheap and simple to perform, they make
accurate material characterization possible, especially for failure analysis and remaining life assessment of in-service
components or structural parts. Nevertheless, a few disadvantages have to be taken into consideration. The small sample
size might not represent the bulk material; the sample size effect influences the material properties; and the results of the
SPT are sensitive to test parameters. However, also for SPT data, ML based models are popular for material parameter
prediction. Most commonly found are ANNs, especially in combination with FEM for data generation. No application of
traditional ML models to SPT data was found in the literature. The paper concludes with an application example which
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
16 of 21 STOLL and BENNER

uses FDCs of structural materials as the basis for predicting he UTS. Simple ML approaches presented here, such as linear
regression models or RFs provide good results for predicting the UTS based on SPT data, even for a very small database.
As a consequence, it is possible to confirm the benefit of simple ML techniques in predicting mechanical properties
such as the UTS based on as simple material tests as the SPT and the authors are sure that ML will positively shape
materials science for the years to come.

ACKNOWLEDGEMENTS
This work is part of the Fraunhofer Lighthouse Project ML4P (Machine Learning for Production). The work profited from
BiGmax, the Max Planck Society’s Research Network on Big-Data-Driven Materials-Science. The data from Section 3 was
kindly provided by Dr. Altstadt of the Helmholtz-Zentrum Dresden-Rossendorf, Germany. Open access funding enabled
and organized by Projekt DEAL.

CONFLICT OF INTEREST
The authors declare no conflicts of interest.

REFERENCES
[1] F. Abbassi, T. Belhadj, S. Mistou, and A. Zghal, Parameter identification of a mechanical ductile damage using artificial neural networks
in sheet metal forming, Mater. Des. 45 (2013), 605–615.
[2] J. A. Abdalla and R. Hawileh, Modeling and simulation of low-cycle fatigue life of steel reinforcing bars using artificial neural network,
J. Frankl. Inst. 348 (2011), 1393–1403.
[3] M. Abendroth and M. Kuna, Determination of deformation and failure properties of ductile materials by means of the small punch test
and neural networks, Comput. Mater. Sci. 28 (2003), 633–644.
[4] M. Abendroth and M. Kuna, Identification of ductile damage and fracture parameters from the small punch test using neural networks,
Eng. Fract. Mech. 73 (2006), 710–725.
[5] M. Abendroth and S. Soltysiak, Assessment of material properties by means of the small punch test, in Recent Trends in Fracture and
Damage Mechanics, Springer International Publishing, Berlin, Germany, 2015, 127–157.
[6] V. S. Agarwala, P. L. Reed, and S. Ahmad, Corrosion detection and monitoring-a review, CORROSION 2000, 2000.
[7] A. Agrawal and A. Choudhary, Perspective: Materials informatics and big data: Realization of the "fourth paradigm" of science in
materials science, Appl. Mater. 4 (2016), 053208.
[8] A. Agrawal, P. D. Deshpande, A. Cecen, G. P. Basavarsu, A. N. Choudhary, and S. R. Kalidindi, Exploration of data science techniques
to predict fatigue strength of steel from composition and processing parameters, Integr. Mater. Manuf. Innov. 3 (2014), 90–108.
[9] E. Altstadt, M. Houska, I. Simonovski, M. Bruchhausen, S. Holmström, and R. Lacalle, On the estimation of ultimate tensile stress from
small punch testing, Int. J. Mech. Sci. 136 (2018), 85–93.
[10] S. Arunkumar, Overview of small punch test, Met. Mater. Int. 26 (2019), 719–738.
[11] S. Arunkumar and R. V. Prakash, Estimation of tensile properties of pressure vessel steel through automated ball indentation and small
punch test, Trans. Indian Inst. Metals 69 (2015), 1245–1256.
[12] D. J. Atha and M. R. Jahanshahi, Evaluation of deep learning approaches based on convolutional neural networks for corrosion detection,
Struct. Health Monit. 17 (2017), 1110–1128.
[13] B. T. Bastian, J. N, S. K. Ranjith, and C. Jiji, Visual inspection and characterization of external corrosion in pipelines using deep neural
network, NDT&E Int. 107 (2019), 102134.
[14] B. Blaiszik, K. Chard, J. Pruyne, R. Ananthakrishnan, S. Tuecke, and I. Foster, The materials data facility: Data services to advance
materials science research, JOM 68 (2016), 2045–2052.
[15] F. E. Bock, R. C. Aydin, C. J. Cyron, N. Huber, S. R. Kalidindi, and B. Klusemann, A review of the application of machine learning and
data mining approaches in continuum materials mechanics, Front Mater 6 (2019), 110.
[16] G. Bradski and A. Kaehler, Learning OpenCV: Computer vision with the OpenCV library, O’Reilly Media, Inc, Sebastopol, CA, 2008.
[17] M. Bruchhausen, S. Holmström, I. Simonovski, T. Austin, J. M. Lapetite, S. Ripplinger, and F. de Haan, Recent developments in small
punch testing: Tensile properties and DBTT, Theor. Appl. Fract. Mech. 86 (2016), 2–10.
[18] J. Bulloch, Toughness losses in low alloy steels at high temperatures: An appraisal of certain factors concerning the small punch test,
Int. J. Press. Vessel. Pip. 75 (1998), 791–804.
[19] K. T. Butler, D. W. Davies, H. Cartwright, O. Isayev, and A. Walsh, Machine learning for molecular and materials science, Nature 559
(2018), 547–555.
[20] T. Byun, E. Lee, J. Hunn, K. Farrell, and L. Mansur, Characterization of plastic deformation in a disk bend test, J. Nucl. Mater. 294 (2001),
256–266.
[21] E. Campitelli, P. Spätig, R. Bonadé, W. Hoffelner, and M. Victoria, Assessment of the constitutive properties from small ball punch test:
Experiment and modeling, J. Nucl. Mater. 335 (2004), 366–378.
[22] CEN CWA-15627, 15627 Worskshop Agreement: Small punch test method for metallic materials, 2006.
[23] J. Chakrabarty, A theory of stretch forming over hemispherical punch heads, Int. J. Mech. Sci. 12 (1970), 315–325.
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
STOLL and BENNER 17 of 21

[24] M. W. Chang, L. Ratinov, and D. Roth, Guiding semi-supervision with constraint-driven learning, Proceedings of the 45th Annual Meeting
of the Association of Computational Linguistics, 2007, pp. 280–287.
[25] S. Chatzidakis, M. Alamaniotis, and L. H. Tsoukalas, Creep rupture forecasting, IJMSTR 2 (2014), 1–25.
[26] L. Christopher, T. Sasikumar, C. Santulli, and C. Fragassa, Neural network prediction of aluminum-silicon carbide tensile strength from
acoustic emission rise angle data, FME Trans. 46 (2018), 253–258.
[27] J. Correa-Baena, K. Hippalgaonkar, J. van Duren, S. Jaffer, V. R. Chandrasekhar, V. Stevanovic, C. Wadia, S. Guha, and T. Buonassisi,
Accelerating materials development via automation, machine learning, and high-performance computing, Joule 2 (2018), 1410–1420.
[28] S. Datta, F. Pettersson, S. Ganguly, H. Saxén, and N. Chakraborti, Designing high strength multi-phase steel for improved
strength-ductility balance using neural networks and multi-objective genetic algorithms, ISIJ Int. 47 (2007), 1195–1203.
[29] M. de Jong, W. Chen, R. Notestine, K. Persson, G. Ceder, A. Jain, M. Asta, and A. Gamst, A statistical learning framework for materials
science: Application to elastic moduli of k-nary inorganic polycrystalline compounds, Sci. Rep. 6 (2016), 34256.
[30] R. K. Desu, S. C. Guntuku, B. Aditya, and A. K. Gupta, Support vector regression based flow stress prediction in austenitic stainless steel
304, Procedia Mater. Sci. 6 (2014), 368–375.
[31] H. Dette, N. Neumeyer, and K. Pilz, A simple nonparametric estimator of a strictly monotone regression function, Bernoulli 12 (2006),
469–490.
[32] P. Dey, J. Bible, S. Datta, S. Broderick, J. Jasinski, M. Sunkara, M. Menon, and K. Rajan, Informatics-aided bandgap engineering for solar
materials, Comput. Mater. Sci. 83 (2014), 185–195.
[33] D. M. Dimiduk, E. A. Holm, and S. R. Niezgoda, Perspectives on the impact of machine learning, deep learning, and artificial intelligence
on materials, processes, and structures engineering, Integr. Mater. Manuf. Innov. 7 (2018), 157–172.
[34] C. Draxl and M. Scheffler, NOMAD: The FAIR concept for big data-driven materials science, MRS Bull. 43 (2018), 676–682.
[35] P. Dymáček, F. Dobeš, P. Král, and J. Dvǒrák, Investigation of fracture in pure aluminium after ECAP by means of small punch test,
Proceedings of the Acta Metallurgica Slovaca-Conference, Vol. 3, 2013, pp. 57–64.
[36] P. Dymáček and K. Milička, Creep small-punch testing and its numerical simulations, Mater. Sci. Eng. A 510-511 (2009), 444–449.
[37] D. Erhan, Y. Bengio, A. Courville, and P. Vincent, Visualizing higher-layer features of a deep network, Univ. Montreal 1341 (2009), 1.
[38] S. Ermon, R. Bras, S. Suram, J. Gregoire, C. Gomes, B. Selman, and R. Van Dover, Pattern decomposition with complex combinatorial
constraints: Application to materials discovery, 2014. arXiv preprint arXiv:1411.7441.
[39] G. G. F. H. Allen and R. Sievers, Crystallographic databases: Information content, software systems, scientific applications, International
Union of Crystallography. Data Commission, Chester, UK, 1987.
[40] J. Fan, F. Han, and H. Liu, Challenges of big data analysis, Natl. Sci. Rev. 1 (2014), 293–314.
[41] S. Fang, M. Wang, W. Qi, and F. Zheng, Hybrid genetic algorithms and support vector regression in forecasting atmospheric corrosion
of metallic materials, Comput. Mater. Sci. 44 (2008), 647–655.
[42] J. Field and M. Swain, A simple predictive model for spherical indentation, J. Mater. Res. 8 (1993), 297–306.
[43] E. Fleury and J. Ha, Small punch tests to estimate the mechanical properties of steels for steam power plant: I. Mechanical strength, Int.
J. Press. Vessel. Pip. 75 (1998), 699–706.
[44] R. Fong and A. Vedaldi, Interpretable explanations of black boxes by meaningful perturbation, Proceedings of the IEEE International
Conference on Computer Vision, 2017, pp. 3429–3437.
[45] C. Fragassa, M. Babic, C. P. Bergmann, and G. Minak, Predicting the tensile behaviour of cast alloys by a pattern recognition analysis
on experimental data, Metals 9 (2019), 557.
[46] O. Frolova, E. Roos, K. Maile, and W. Müller, Representation of the heat specific creep rupture behaviour of 9% Cr steels using neural
networks, Trans. MLDM 4 (2011), 1–16.
[47] M. W. Gaultois, A. O. Oliynyk, A. Mar, T. D. Sparks, G. J. Mulholland, and B. Meredig, Perspective: Web-based machine learning models
for real-time screening of thermoelectric materials properties, APL Mater. 4 (2016), 053213.
[48] B. Gülçimen and P. Hähner, Determination of creep properties of a p91 weldment by small punch testing and a new evaluation approach,
Mater. Sci. Eng. A 588 (2013), 125–131.
[49] J. Graser, S. K. Kauwe, and T. D. Sparks, Machine learning and energy minimization approaches for crystal structure predictions: A
review and new horizons, Chem. Mater. 30 (2018), 3601–3612.
[50] C. R. Groom, I. J. Bruno, M. P. Lightfoot, and S. C. Ward, The Cambridge structural database, Acta Crystallogr. Sect. B 72 (2016), 171–179.
[51] A. K. Gupta, H. N. Krishnamurthy, Y. Singh, K. M. Prasad, and S. K. Singh, Development of constitutive models for dynamic strain
aging regime in austenitic stainless steel 304, Mater. Des. 45 (2013), 616–627.
[52] G. Gutscher, H. C. Wu, G. Ngaile, and T. Altan, Determination of flow stress for sheet metal forming using the viscous pressure bulge
(VPB) test, J. Mater. Process. Technol. 146 (2004), 1–7.
[53] C. Heintze, C. Recknagel, F. Bergner, M. Hernández-Mayoral, and A. Kolitsch, Ion-irradiation-induced damage of steels characterized by
means of nanoindentation, Nucl. Instrum. Methods Phy. Res. B 267 (2009), 1505–1508 Proceedings of the 16th International Conference
on Ion Beam Modification of Materials.
[54] N. D. Hoang and V. D. Tran, Image processing-based detection of pipe corrosion using texture analysis and metaheuristic-optimized
machine learning approach, Comput. Intell. Neurosci. 2019 (2019), 1–13.
[55] M. Houska and E. Altstadt, SP tests on p92 (T70175) at RT, Eur Commiss JRC SP tests on P92 (T70175) at RT Version 1.0, 2017.
[56] N. Huber, A. Konstantinidis, and C. Tsakmakis, Determination of Poisson’s ratio by spherical indentation using neural networks—Part
I: Theory, J. Appl. Mech. 68 (2000), 218–223.
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
18 of 21 STOLL and BENNER

[57] N. Huber, D. Munz, and C. Tsakmakis, Determination of Young’s modulus by spherical indentation, J. Mater. Res. 12 (1997),
2459–2469.
[58] R. Hurst and K. Matocha, The european code of practice for small punch testing – Where do we go from here, Metall. J. 63 (2010), 5–11.
[59] R. Hurst and K. Matocha, Where are we now with the European code of practice for small punch testing, in Determination Mechanical
Properties Materials by Small Punch Other Miniature Testing Techniques, K. Matocha, R. Hurst, and W. Sun, Eds., OCELOT SRO, Ostrava,
2012, 4–18.
[60] A. Husain, R. Sharma, and D. Sehgal, Small punch and indentation tests for structural health monitoring, Process. Eng. 173 (2017),
710–717.
[61] O. Isayev, D. Fourches, E. N. Muratov, C. Oses, K. Rasch, A. Tropsha, and S. Curtarolo, Materials cartography: Representing and mining
materials space using structural and electronic fingerprints, Chem. Mater. 27 (2015), 735–743.
[62] A. Jain, G. Hautier, S. Ong, and K. Persson, New opportunities for materials informatics: Resources and data mining techniques for
uncovering hidden relationships, J. Mater. Res. 31 (2016), 977–994.
[63] A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson,
Commentary: The materials project: A materials genome approach to accelerating materials innovation, APL Mater. 1 (2013), 011002.
[64] G. James, D. Witten, T. Hastie, and R. Tibshirani, An introduction to statistical learning, Vol 112, Springer, New York, NY, 2013.
[65] A. Janča, J. Siegl, and P. Haušild, Small punch test evaluation methods for material characterisation, J. Nucl. Mater. 481 (2016), 201–213.
[66] M. Jiménez-Come, I. Turias, and F. Trujillo, An automatic pitting corrosion detection approach for 316L stainless steel, Mater. Des. 56
(2014), 642–648.
[67] I. T. Jolliffe and J. Cadima, Principal component analysis: A review and recent developments, Philos. Trans. R. Soc. A 374 (2016),
20150202.
[68] M. Jordan and T. Mitchell, Machine learning: Trends, perspectives, and prospects, Science 349 (2015), 255–260.
[69] J. Kameda, A kinetic model for ductile-brittle fracture mode transition behavior, Acta Metall. 34 (1986), 2391–2398.
[70] J. Kameda and X. Mao, Small-punch and TEM-disc testing techniques and their application to characterization of radiation damage,
J. Mater. Sci. 27 (1992), 983–989.
[71] E. Kim, K. Huang, S. Jegelka, and E. Olivetti, Virtual screening of inorganic materials synthesis parameters with deep learning, NPJ
Comput. Mater. 3 (2017), 1–9.
[72] S. W. Kim and Y. S. Lee, Comparative study on failure prediction in warm forming processes of Mg alloy sheet by the FEM and ductile
fracture criteria, Metall. Mater. Trans. B: Process Metall. Mater. Process. Sci. 45 (2013), 445–453.
[73] S. Kirklin, J. E. Saal, B. Meredig, A. Thompson, J. W. Doak, M. Aykol, S. Rühl, and C. Wolverton, The open quantum materials database
(OQMD): Assessing the accuracy of DFT formation energies, NPJ Comput. Mater. 1 (2015), 1–15.
[74] I. Klevtsov, A. Dedov, and A. Molodtsov, Using of small punch test for determination of tensile properties for power plant steels, Proceedings
of the 6th International DAAAM Baltic Conference Industrial Engineering, 2008.
[75] S. Kohlar, Gefüge und eigenschaften des warmfesten chromstahls p91, Wissenschaftlich-Technische Berichte/Helmholtz-Zentrum
Dresden-Rossendor HZDR-082 2017, 2017.
[76] P. Konopík, J. Dzugan, and R. Prochazka, Determination of fracture toughness and tensile properties of structural steels by small punch
test and micro-tensile test, Metal Brno Czech Republic 2013 (2013), 15–17.
[77] K. Kumar, A. Pooleery, K. Madhusoodanan, R. Singh, J. Chakravartty, R. Shriwastaw, B. Dutta, and R. Sinha, Evaluation of ultimate
tensile strength using miniature disk bend test, J. Nucl. Mater. 461 (2015), 100–111.
[78] S. H. Kwon, D. G. Hong, and C. H. Yim, Prediction of hot ductility of steels from elemental composition and thermal history by deep
neural networks, Ironmak. Steelmak. 47 (2020), 1176–1187.
[79] F. Legrain, J. Carrete, A. van Roekeghem, S. Curtarolo, and N. Mingo, How chemical composition alone can predict vibrational free
energies and entropies of solids, Chem. Mater. 29 (2017), 6220–6227.
[80] Z. Li, J. R. Kermode, and A. De Vita, Molecular dynamics with on-the-fly machine learning of quantum-mechanical forces, Phys. Rev.
Lett. 114 (2015), 096405.
[81] Z. Li, S. Wang, W. S. Chin, L. E. Achenie, and H. Xin, High-throughput screening of bimetallic catalysts enabled by machine learning,
J. Mater. Chem. A 5 (2017), 24131–24138.
[82] Y. Lin, J. Zhang, and J. Zhong, Application of neural networks to predict the elevated temperature flow behavior of a low alloy steel,
Comput. Mater. Sci. 43 (2008), 752–758.
[83] T. Linse, M. Kuna, J. Schuhknecht, and H. W. Viehrig, Usage of the small-punch-test for the characterisation of reactor vessel steels in
the brittle-ductile transition region, Eng. Fract. Mech. 75 (2008), 3520–3533.
[84] T. Linse, M. Kuna, and H. W. Viehrig, Quantification of brittle-ductile failure behavior of ferritic reactor pressure vessel steels using the
small-punch-test and micromechanical damage models, Mater. Sci. Eng. A 614 (2014), 136–147.
[85] Q. Liu, X. Zhang, B. Wang, and B. Wang, Control technology of solidification and cooling in the process of continuous casting of steel,
Sci. Technol. Cast. Process. 26 (2012), 169–204.
[86] Y. Liu, T. Zhao, W. Ju, and S. Shi, Materials discovery and design using machine learning, J. Mater. 3 (2017), 159–177.
[87] L. Lu, M. Dao, P. Kumar, U. Ramamurty, G. Karniadakis, and S. Suresh, Extraction of mechanical properties of materials through deep
learning from instrumented indentation, Proc. Natl. Acad. Sci. U. S. A. 117 (2020), 7052–7062.
[88] A. Lucas, Connecting microscopic physics with the macroscopic properties of materials in introductory physics courses, 2014, arXiv preprint
arXiv:1402.2593.
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
STOLL and BENNER 19 of 21

[89] X. Ma, Z. Li, L. E. Achenie, and H. Xin, Machine-learning-augmented chemisorption model for CO2 electroreduction catalyst screening,
J. Phys. Chem. Lett. 6 (2015), 3528–3533.
[90] J. MacInnes, S. Santosa, and W. Wright, Visual classification: Expert knowledge guides machine learning, IEEE Comput. Graph. Appl.
30 (2010), 8–14.
[91] G. Mahalle, O. Salunke, N. Kotkunde, A. K. Gupta, and S. K. Singh, Neural network modeling for anisotropic mechanical properties
and work hardening behavior of inconel 718 alloy at elevated temperatures, J. Mater. Res. Technol. 8 (2019), 2130–2140.
[92] M. Manahan, A. Argon, and O. Harling, The development of a miniaturized disk bend test for the determination of postirradiation
mechanical properties, J. Nucl. Mater. 104 (1981), 1545–1550.
[93] M. P. Manahan, The development of a miniaturized disk bend test for the determination of post-irradiation mechanical behavior, Ph.D.
Thesis, Massachusetts Institute of Technology, 1982.
[94] A. Mannodi-Kanakkithodi, T. D. Huan, and R. Ramprasad, Mining materials design rules from data: The example of polymer dielectrics,
Chem. Mater. 29 (2017), 9001–9010.
[95] X. Mao, M. Saito, and H. Takahashi, Small punch test to predict ductile fracture toughness JIC and brittle fracture toughness KIC, Scr.
Metall. Mater. 25 (1991), 2481–2485.
[96] X. Mao and H. Takahashi, Development of a further-miniaturized specimen of 3 mm diameter for tem disk (ø 3 mm) small punch tests,
J. Nucl. Mater. 150 (1987), 42–52.
[97] Materials Information Science National Institute of Materials Science, SuperCon, https://supercon.nims.go.jp/en/. Accessed July 21,
2020.
[98] Materials Project, https://materialsproject.org/. Accessed July 21, 2020.
[99] J. McNaney, G. Lucas, and G. Odette, Application of ball punch tests to evaluating fracture mode transition in ferritic steels, J. Nucl.
Mater. 179-181 (1991), 429–433.
[100] B. Medasani, A. Gamst, H. Ding, W. Chen, K. A. Persson, M. Asta, A. Canning, and M. Haranczyk, Predicting defect behavior in B2
intermetallics by merging ab initio modeling and machine learning, NPJ Comput. Mater. 2 (2016), 1–10.
[101] L. Meng, P. Breitkopf, B. Raghavan, G. Mauvoisin, O. Bartier, and X. Hernot, Identification of material properties using indentation test
and shape manifold learning approach, Comput. Methods Appl. Mech. Eng. 297 (2015), 239–257.
[102] L. Meng, B. Raghavan, O. Bartier, X. Hernot, G. Mauvoisin, and P. Breitkopf, An objective meta-modeling approach for indentation-based
material characterization, Mech. Mater. 107 (2017), 31–44.
[103] E. Metzbower, J. deLoach, S. Lalam, and H. Bhadeshia, Neural network analysis of strength and ductility of welding alloys for high
strength low alloy shipbuilding steels, Sci. Technol. Weld. Join. 6 (2001), 116–124.
[104] M. A. Meyers and K. K. Chawla, Mechanical behavior of materials, Cambridge University Press, Cambridge, MA, 2008.
[105] K. Milička and F. Dobeš, Small punch testing of p91 steel, Int. J. Press. Vessel. Pip. 83 (2006), 625–634.
[106] T. Misawa, T. Adachi, M. Saito, and Y. Hamaguchi, Small punch tests for evaluating ductile-brittle transition behavior of irradiated
ferritic steels, J. Nucl. Mater. 150 (1987), 194–202.
[107] T. Misawa, S. Nagata, N. Aoki, J. Ishizaka, and Y. Hamaguchi, Fracture toughness evaluation of fusion reactor structural steels at low
temperatures by small punch tests, J. Nucl. Mater. 169 (1989), 225–232.
[108] A. Mlinarić, M. Horvat, and V. Šupak Smolčić, Dealing with the positive publication bias: Why you should really publish your negative
results, Biochem. Med. 27 (2017), 447–452.
[109] T. Mueller, A. G. Kusne, and R. Ramprasad, Machine learning in materials science: Recent progress and emerging applications, in Reviews
in Computational Chemistry, A. Parrill and K. Lipkowitz, Eds., Wiley Online Library, New Jersey, USA, 2016, 186–273.
[110] National Academy of Sciences (US). Committee on the Survey of Materials Science, National Academy of Sciences (US). Committee on
Science, and Public Policy, Materials and Man’s Needs: Materials Science and Engineering; Supplementary Report, Vol. 1, National
Academies, 1974.
[111] National Institute of Materials Science, https://www.nims.go.jp/eng/. Accessed July 21, 2020.
[112] NIST Materials Data Repository. https://materialsdata.nist.gov/. Accessed July 21, 2020.
[113] Nomad Repository. https://www.nims.go.jp/eng/. Accessed July 21, 2020.
[114] W. Oliver and G. Pharr, An improved technique for determining hardness and elastic modulus using load and displacement sensing
indentation experiments, J. Mater. Res. 7 (1992), 1564–1583.
[115] G. Olson, Computational design of hierarchically structured materials, Science 277 (1997), 1237–1242.
[116] J. O’Mara, B. Meredig, and K. Michel, Materials data infrastructure: A case study of the citrination platform to examine data import,
storage, and access, JOM 68 (2016), 2031–2034.
[117] S. Pattanayak, S. Dey, S. Chatterjee, S. G. Chowdhury, and S. Datta, Computational intelligence based designing of microalloyed pipeline
steel, Comput. Mater. Sci. 104 (2015), 60–68.
[118] I. Peñuelas, C. Betegón, C. Rodríguez, and J. Belzunce, Inverse methods on small punch tests, Numerical Simulations-Applications,
Examples Theory, INCTECH, Rijeka (Croatia), 2011, pp. 311–330.
[119] G. Pilania, A. Mannodi-Kanakkithodi, B. Uberuaga, R. Ramprasad, J. Gubernatis, and T. Lookman, Machine learning bandgaps of
double perovskites, Sci. Rep. 6 (2016), 19375.
[120] R. C. Poudel, T. Sakaguchi, and Y. Shimizu, A selective approach on data based quality prediction for quenched and tempered steel
reinforcement bars, J. Chem. Eng. Jpn 46 (2013), 294–301.
[121] B. Puchala, G. Tarcea, E. A. Marquis, M. Hedstrom, H. Jagadish, and J. E. Allison, The materials commons: A collaboration platform
and information repository for the global materials community, JOM 68 (2016), 2035–2044.
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
20 of 21 STOLL and BENNER

[122] P. Raccuglia, K. C. Elbert, P. D. Adler, C. Falk, M. B. Wenny, A. Mollo, M. Zeller, S. A. Friedler, J. Schrier, and A. J. Norquist,
Machine-learning-assisted materials discovery using failed experiments, Nature 533 (2016), 73–76.
[123] R. Rahman and S. Zhafer Firdaus Syed Putra, Tensile properties of natural and synthetic fiber-reinforced polymer composites, in Mechanical
and Physical Testing of Biocomposites, Fibre-Reinforced Composites and Hybrid Composites, Elsevier, Sawston, UK, 2019, 81–102.
[124] M. Raissi, P. Perdikaris, and G. Karniadakis, Physics informed deep learning (part I): Data-driven solutions of nonlinear partial differential
equations, 2017. arXiv preprint arXiv:1711.10561.
[125] K. Rajan, Materials informatics, Mater. Today 8 (2005), 38–45.
[126] S. Rühl, The inorganic crystal structure database (ICSD): A tool for materials sciences. https://icsd.products.fiz-karlsruhe.de/. Accessed
July 21, 2020.
[127] J. E. Saal, S. Kirklin, M. Aykol, B. Meredig, and C. Wolverton, Materials design and discovery with high-throughput density functional
theory: The open quantum materials database (OQMD), JOM 65 (2013), 1501–1509.
[128] C. Sainte Catherine, J. Messier, C. Poussard, S. Rosinski, and J. Foulds, Small punch test: Epri-cea finite element simulation benchmark
an inverse method for the estimation of elastic plastic behavior, Am. Soc. Test. Mater. 4 (2002), 350–370.
[129] T. Sangkharat and S. Dechjarern, Using image processing on erichsen cup test machine to calculate anisotropic property of sheet metal,
Proc. Manuf. 29 (2019), 390–397.
[130] I. Sankar, K. Rao, and B. Murhty, Applying ANN to predict mechanical properties in TMT process, IE (I) Journal-MM 90 (2009), 3–6.
[131] I. Santos, J. Nieves, Y. K. Penya, and P. G. Bringas, Machine-learning-based mechanical properties prediction in foundry production,
ICCAS-SICE 2009 (2009), 4536–4541.
[132] T. Sasikumar, S. Rajendraboopathy, K. Usha, and E. Vasudev, Artificial neural network prediction of ultimate strength of unidirectional
T-300/914 tensile specimens using acoustic emission response, J. Nondestruct. Eval. 27 (2008), 127–133.
[133] J. Schmidt, M. R. Marques, S. Botti, and M. A. Marques, Recent advances and applications of machine learning in solid-state materials
science, NPJ Comput. Mater. 5 (2019), 1–36.
[134] K. Schütt, H. Glawe, F. Brockherde, A. Sanna, K. Müller, and E. Gross, How to represent crystal structures for machine learning: Towards
fast prediction of electronic properties, Phys. Rev. B 89 (2014), 205118.
[135] A. Seko, H. Hayashi, K. Nakayama, A. Takahashi, and I. Tanaka, Representation of compounds for machine-learning prediction of
physical properties, Phys. Rev. B 95 (2017), 144110.
[136] C. Shen, C. Wang, X. Wei, Y. Li, S. van der Zwaag, and W. Xu, Physical metallurgy-guided machine learning and artificial intelligent
design of ultrahigh-strength stainless steel, Acta Mater. 179 (2019), 201–214.
[137] H. Shigemori, M. Kano, and S. Hasebe, Optimum quality design system for steel products through locally weighted regression model,
J. Process Control 21 (2011), 293–301.
[138] H. Shigemori and S. Kawamura, Optimum quality design support system for steel products using locally-weighted regression model,
Proceedings of the SICE Annual Conference 2007, September, IEEE, 2007, pp. 810–815.
[139] D. Shin, Y. Yamamoto, M. P. Brady, S. Lee, and J. Haynes, Modern data analytics approach to predict creep of high-temperature alloys,
SSRN J. 168 (2018), 321–330.
[140] T. Shiraiwa, F. Briffod, Y. Miyazawa, and M. Enoki, Fatigue performance prediction of structural materials by multi-scale modeling and
machine learning, Proceedings of the 4th World Congress on Integrated Computational Materials Engineering (ICME 2017), 2017,
pp. 317–326.
[141] I. Simonovski, S. Holmström, and M. Bruchhausen, Small punch tensile testing of curved specimens: Finite element analysis and
experiment, Int. J. Mech. Sci. 120 (2017), 204–213.
[142] K. Simonyan, A. Vedaldi, and A. Zisserman, Deep inside convolutional networks: Visualising image classification models and saliency
maps, 2013. arXiv preprint arXiv:1312.6034.
[143] Small punch test methods of metallic materials for in-service pressure equipment, GB/T 29459, https://www.chinesestandard.net/PDF.
aspx/GBT29459.2-2012. Accessed July 21, 2020.
[144] M. Song, K. Guan, W. Qin, and J. A. Szpunar, Comparison of mechanical properties in conventional and small punch tests of fractured
anisotropic A350 alloy forging flange, Nucl. Eng. Des. 247 (2012), 58–65.
[145] Q. Song, A preliminary investigation on materials informatics, Chin. Sci. Bull. 49 (2004), 210.
[146] T. Sourmail, H. Bhadeshia, and D. MacKay, Neural network model of creep strength of austenitic stainless steels, Mater. Sci. Technol.
18 (2002), 655–663.
[147] Standard Method for Small Punch Testing of Ultra-High Molecular Weight Polyethylene Used in Surgical Implants, ASTM F2183-02,
2008, Accessed July 21, 2020.
[148] Z. Sterjovski, D. Nolan, K. Carpenter, D. Dunne, and J. Norrish, Artificial neural networks for modelling the mechanical properties of
steels in various applications, J. Mater. Process. Technol. 170 (2005), 536–544.
[149] J. Struyf and S. Džeroski, Clustering trees with instance level constraints, Proceedings of the European Conference on Machine Learning,
2007, pp. 359–370.
[150] M. Sundararajan, A. Taly, and Q. Yan, Axiomatic attribution for deep networks, 2017. arXiv preprint arXiv:1703.01365.
[151] S. Swaddiwudhipong, K. K. Tho, Z. S. Liu, J. Hua, and N. S. B. Ooi, Material characterization via least squares support vector machines,
Model. Simul. Mater. Sci. Eng. 13 (2005), 993–1004.
[152] D. Tabor, The hardness and strength of metals, J. Inst. Met. 79 (1951), 1.
[153] B. Taljat, T. Zacharia, and F. Haggag, Analysis of ball-indentation load-depth data: Part I. determining elastic modulus, J. Mater. Res. 12
(1997), 965–974.
 15222608, 2021, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/gamm.202100003 by FISH & RICHARDSON, Wiley Online Library on [04/12/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
STOLL and BENNER 21 of 21

[154] A. A. Tavassoli, A. Alamo, L. Bedel, L. Forest, J. M. Gentzbittel, J. W. Rensman, E. Diegele, R. Lindau, M. Schirra, R. Schmitt, H. Schnei-
der, C. Petersen, A. M. Lancha, P. Fernandez, G. Filacchioni, M. Maday, K. Mergia, N. Boukos, P. Baluc, E. Alves Spätig, and E. Lucon,
Materials design data for reduced activation martensitic steel type EUROFER, J. Nucl. Mater. 329-333 (2004), 257–262 Proceedings of
the 11th International Conference on Fusion Reactor Materials (ICFRM-11).
[155] The Materials Genome Initiative, https://www.mgi.gov/. Accessed July 21, 2020.
[156] Topological Materials Database, https://www.topologicalquantumchemistry.com/. Accessed September 21, 2020.
[157] Z. W. Ulissi, A. J. Medford, T. Bligaard, and J. K. Nørskov, To address surface reaction network complexity using scaling relations machine
learning and DFT calculations, Nat. Commun. 8 (2017), 1–7.
[158] A. Vellido, J. D. Martín-Guerrero, and P. J. Lisboa, Making machine learning models interpretable, ESANN 12 (2012), 163–172.
[159] M. Verleysen and D. François, The curse of dimensionality in data mining and time series prediction, Proceedings of the International
Work-Conference on Artificial Neural Networks, 2005, pp. 758–770.
[160] H. W. Viehrig, M. Scibetta, and K. Wallin, Application of advanced master curve approaches on WWER-440 reactor pressure vessel
steels, Int. J. Press. Vessel. Pip. 83 (2006), 584–592.
[161] L. von Rueden, S. Mayer, K. Beckh, B. Georgiev, S. Giesselbach, R. Heese, B. Kirsch, J. Pfrommer, A. Pick, and R. Ramamurthy, Informed
machine learning – A taxonomy and survey of integrating knowledge into learning systems, 2020. arXiv preprint arXiv:1903.12394.
[162] N. Wagner and J. M. Rondinelli, Theory-guided machine learning in materials science, Front. Mater. 3 (2016), 28.
[163] C. Wang, C. Shen, Q. Cui, C. Zhang, and W. Xu, Tensile property prediction by feature engineering guided machine learning in reduced
activation ferritic/martensitic steels, J. Nucl. Mater. 529 (2020), 151823.
[164] C. Wang, C. Shen, X. Huo, C. Zhang, and W. Xu, Design of comprehensive mechanical properties by machine learning and
high-throughput optimization algorithm in RAFM steels, Nucl. Eng. Technol. 52 (2020), 1008–1012.
[165] H. Wang, T. Xu, and B. Shou, Determination of material strengths by hydraulic bulge test, Materials 10 (2016), 23.
[166] Z. X. Wang, H. J. Shi, J. Lu, P. Shi, and X. F. Ma, Small punch testing for assessing the fracture properties of the reactor vessel steel with
different thicknesses, Nucl. Eng. Des. 238 (2008), 3186–3193.
[167] L. Ward, M. Aykol, B. Blaiszik, I. Foster, B. Meredig, J. Saal, and S. Suram, Strategies for accelerating the adoption of materials informatics,
MRS Bull. 43 (2018), 683–689.
[168] J. Wei, X. Chu, X. Y. Sun, K. Xu, H. X. Deng, J. Chen, Z. Wei, and M. Lei, Machine learning in materials science, InfoMat 1 (2019),
338–358.
[169] Q. Wei, X. Peng, X. Liu, and W. Xie, Materials informatics and study on its further development, Chin. Sci. Bull. 51 (2006), 498–504.
[170] Y. Wen, C. Cai, X. Liu, J. Pei, X. Zhu, and T. Xiao, Corrosion rate prediction of 3C steel under different seawater environment by using
support vector regression, Corros. Sci. 51 (2009), 349–355.
[171] A. White, The materials genome initiative: One year on, MRS Bull. 37 (2012), 715–716.
[172] S. Wold, K. Esbensen, and P. Geladi, Principal component analysis, Chemom. Intell. Lab. Syst. 2 (1987), 37–52.
[173] R. N. Wright, Wire technology, Elsevier, Amsterdam, Netherlands, 2016.
[174] M. Yetna N’Jock, D. Chicot, X. Decoopman, J. Lesage, J. Ndjaka, and A. Pertuz, Mechanical tensile properties by spherical macroinden-
tation using an indentation strain-hardening exponent, Int. J. Mech. Sci. 75 (2013), 257–264.
[175] A. Zakutayev, N. Wunder, M. Schwarting, J. D. Perkins, R. White, K. Munch, W. Tumas, and C. Phillips, An open experimental database
for exploring inorganic materials, Sci Data 5 (2018), 180053.
[176] M. Zeiler and R. Fergus, Visualizing and understanding convolutional networks, Proceedings of the European Conference on Computer
Vision, 2014, pp. 818–833.
[177] E. Zhang, M. Yin, and G. Karniadakis, Physics-informed neural networks for nonhomogeneous material identification in elasticity imaging,
2020. arXiv preprint arXiv:2009.04525.
[178] L. Zhang, J. Lei, Q. Zhou, and Y. Wang, Using genetic algorithm to optimize parameters of support vector machine and its application
in material fatigue life prediction, Adv. Nat. Sci. 8 (2015), 21–26.
[179] Y. Zhang and C. Ling, A strategy to apply machine learning to small datasets in materials science, NPJ Comput. Mater. 4 (2018), 1–8.
[180] Y. Zhu, N. Zabaras, P. Koutsourelakis, and P. Perdikaris, Physics-constrained deep learning for high-dimensional surrogate modeling
and uncertainty quantification without labeled data, J. Comput. Phys. 394 (2019), 56–81.
[181] C. Zurbuchen, Influence of specimen type, crack length and evaluation method on quasi-static and dynamic fracture toughness properties,
Proceedings of the ASME Pressure Vessels Piping Conference Volume 5: High Pressure Technology; Nondestructive Evaluation Division;
Student Paper Competition, July 2009, pp. 511–517.

How to cite this article: Stoll A, Benner P. Machine learning for material characterization with an application
for predicting mechanical properties. GAMM-Mitteilungen. 2021;44:e202100003. https://doi.org/10.1002/gamm.
202100003