Materials & Design 244 (2024) 113224

Contents lists available at ScienceDirect

Materials & Design

journal homepage: www.elsevier.com/locate/matdes

Few-shot learning-based generative design of metamaterials with zero

Poisson’s ratio
Xiangbei Liu a , Huan Zhao a , Ya Tang a , Chaofan Chen b , Yifeng Zhu c , Bo Song d , Yan Li a, *
a
Thayer School of Engineering, Dartmouth College, Hanover, NH 03755, USA
b
School of Computing and Information Science, University of Maine, Orono, ME 04469, USA
c
Department of Electrical and Computer Engineering, University of Maine, Orono, ME 04469, USA
d
Sandia National Laboratories, 1515 Eubank SE, Albuquerque, NM 87185, USA

A R T I C L E I N F O A B S T R A C T

Keywords: Metamaterials with a zero Poisson’s ratio offer significant advantages in robotic actuation and space exploration
Machine learning due to their precise control of deformation. However, existing machine learning techniques cannot be directly
Few-shot learning used to accelerate the design of such materials due to the scarcity of this property. We propose a few-shot
Metamaterials
learning-based framework to generate non-periodic metamaterials with zero Poisson’s ratio. Our framework
Conditional variational autoencoder (cVAE)
Out-of-distribution (OOD) data
incorporates an out-of-distribution (OOD) target-oriented sampler into a conditional variational autoencoder
Inverse design (cVAE). Unlike other metamaterial generative models that only deal with continuous pixel data, our approach
Possion’s ratio can handle discrete unit cell patterns by computing their probability distributions. We found that controlling the
learning focus during the training process can effectively mitigate the scarcity of acceptable data within the
training set. This mitigation is achieved by repeatedly selecting target samples through the OOD target-oriented
sampler. Incorporating active learning into the training process can further enhance model efficiency by adap-
tively adjusting the ratio between acceptable and unacceptable samples. The impacts of training data size,
effective data composition, and the number of iterations in active learning on design efficiency are discussed in
detail. Compared to random trial-and-error generation, our model demonstrates a substantial increase in the
acceptable rate, from 0.3 % to 39 %.

1. Introduction that these existing designs with a zero Poisson’s ratio metamaterial
mostly involve a periodic arrangement of the same unit cell [8–10].
Materials with a Poisson’s ratio of zero have significant potential for While these designs are straightforward, the structural regularity may
various applications, such as actuators, scaffolds, and gaskets. For limit the range of achievable functionalities [11]. Drawing inspiration
example, the aorta in Fig. 1(a) may experience pressure changes due to from the irregular structure of cork, a natural material known for its zero
the pulsatile flow of blood from the heart. By using a material with a zero Poisson’s ratio [12], incorporating non-periodic designs into the
Poisson’s ratio in the scaffold, stability can be achieved as it prevents development of zero Poisson’s ratio metamaterials presents a promising
expansion or contraction under pressure. This stability helps to prevent approach to address this challenge. It is reported that disordered cellular
excessive deformation of the aorta and reduces the risk of complications networks tend to exhibit a zero Poisson’s ratio in the presence of topo-
such as aneurysm formation or vessel damage. However, traditional logical defects [13] and also be affected by the microscale fluctuations in
engineering materials rarely possess this unique property. In recent the material [5]. However, achieving such designs typically relies on
years, metamaterial research has made noteworthy advancements in parametric studies [14] or topology optimization [15] to customize the
creating novel properties that are not commonly found in bulk materials characteristics of metamaterials [16]. Consequently, these approaches
[1–4]. It has been discovered that materials with reentrant and trian- often suffer from low efficiency as they require repetitive simulations or
gular architectures can exhibit a negative Poisson’s ratio [5]. Through experiments to fine-tune or adjust the search strategy [17]. Moreover,
iterative geometric optimization, several metamaterial designs with a the success of the optimization process heavily depends on the expertise
zero Poisson’s ratio have been developed [6,7]. It is important to note of the designer [18]. Therefore, given the exponential growth in the

* Corresponding author.
E-mail address: [email protected] (Y. Li).

https://doi.org/10.1016/j.matdes.2024.113224
Received 15 May 2024; Received in revised form 8 July 2024; Accepted 30 July 2024
Available online 31 July 2024
0264-1275/© 2024 Dartmouth College. Published by Elsevier Ltd. This is an open access article under the CC BY-NC license
(http://creativecommons.org/licenses/by-nc/4.0/).
X. Liu et al. Materials & Design 244 (2024) 113224

number of potential combinations of substructures, traditional trial-and- metamaterials with specific properties.
error design methods can be both expensive and time-consuming. Nevertheless, despite the advancements in using ML for meta-
Machine learning (ML) presents a promising alternative for gener- material design, the aforementioned approaches cannot be directly
ating metamaterials with specific properties, addressing the aforemen- applied to generate non-periodic metamaterials with a zero Poisson’s
tioned inefficiency and dependence on designer expertise. For instance, ratio. Firstly, due to the scarcity of available data on desired properties,
neural networks (NNs) in combination with various techniques have the training dataset primarily consists of unacceptable metamaterial
been employed in metamaterial designs. Zeng et al. [19] presented a designs. Traditional generative machine learning models, such as GAN
deep multi-NN for customizing metamaterials with a targeted Poisson’s [24] and variational autoencoder (VAE), are unable to handle data with
ratio ranging from − 1 to 1, based on a pixelated metamaterial database out-of-distribution (OOD) targets. Secondly, previous studies on meta-
comprising 3,500 training data points. Alderete et al. [20] introduced a materials have employed either a pixelated geometry [17–19] or some
machine learning framework, involving clustering, deep NN, and ge- parametric geometry representation [20–23]. The former approach may
netic algorithms, to rationally guide the design of the kirigami meta- lead to designs that are impossible to manufacture due to their non-
materials with controlled deformation behavior. This framework connected features, which require an additional supplementary algo-
utilized 20,000 samples defined by four geometric parameters. Tanri- rithm to rectify this issue [17]. The latter approach may restrict the
over et al. [21] utilized an autoencoder and a fully connected NN to models’ scalability to accommodate other dimensions or new geometric
guide the design of free-form dielectric metasurfaces. Their model was features that cannot be supported by the parametrization. Therefore, a
trained using a large database of over 32,000 metasurfaces. Generative new representation format for metamaterials is necessary.
ML models, such as generative adversarial networks (GAN), have also To address these challenges, this work presents an innovative few-
been widely utilized. Challapalli et al. [22] used a combination of GAN shot learning-based framework for metamaterial design. Few-shot
and regression models to discover cellular metamaterials with superior learning, an important subfield of machine learning, offers a solution
strength and recovery stress. Their training dataset included 1,500 lat- to extract the underlying metamaterial patterns from a training set with
tice units formed by representative volume elements and their corre- limited acceptable samples [25–27]. As illustrated in Fig. 1(b), this
sponding mechanical properties. Tian et al. [18] utilized a cycle GAN approach significantly enhances the success rate of metamaterial design.
and a conventional convolutional neural network (CNN) to inversely Details of the framework are elaborated in Section 2. Additionally,
design metamaterials based on a target Poisson’s ratio-strain curve. Section 3 discusses various factors that impact the efficiency of material
Their dataset comprised 2,000 metamaterial structures, represented in a design, including the size of training data, the influence of effective data
200 × 200 pixel format. Zheng et al. [23] developed a conditional GAN composition in OOD target-oriented cVAE, and the importance of
to generate metamaterials with a user-defined Poisson’s ratio and employing an active learning strategy. Through systematic studies, we
Young’s modulus. Their model was trained using a dataset with 100,000 aim to address two critical questions in materials design: (1) How much
metamaterials, each represented by 64 coordinate points. These meth- training data is sufficient for the machine learning model to provide
odologies have demonstrated significant success in designing reliable feedback on material design? (2) Given the constraints of a

Fig. 1. (a) Scheme of the scaffold attached to the aorta, featuring a metamaterial with a zero Poisson’s ratio designed by our model. This metamaterial is integrated
into the scaffold’s wall to resist lateral expansion when subjected to in-plane compression from blood flow, ensuring the scaffold remains stable without damaging the
aorta. (b) Comparison of the design efficiency in finding zero Poisson’s ratio metamaterials using random generation, our ODD target-oriented cVAE model with and
without an active learning strategy. (c) Four unit cell candidates [28,29].

2
X. Liu et al. Materials & Design 244 (2024) 113224

limited number of samples that meet the design criteria, how can the positioned between two steel clamping blocks, and a strain rate of 4.2 ×
model be fine-tuned to shift its learning focus? 10− 4 s− 1 was applied. The reaction force and displacement of the
loading block were measured and recorded with a digital oscilloscope at
2. Methodology a sampling rate of 10 Hz. A Canon EOS 6D camera with a Canon 70–200
mm zoom lens was placed above the sample to capture the in-situ
2.1. Training data generation metamaterial deformation process. The experimental result is illus-
trated in Fig. 3. It is evident from Fig. 3 that both the nominal stress–-
As shown in Fig. 1(c), when creating the non-periodic metamaterial strain curve (Fig. 3(a)) and the deformation behavior (Fig. 3(b)) of the
samples, a pool of four preselected unit cell candidates with 15 mm side metamaterial sample predicted by the finite element simulation
dimensions is used. When arranged in a periodic lattice pattern, the matched well with the experimental results.
associated Poisson’s ratios are as follows: square (ν = 0.087), honey- As illustrated in Fig. 3(a), ten evenly spaced reference lines were
comb (ν = 0.536), reentrant (ν = − 0.272), and triangle (ν = − 0.932). introduced to each metamaterial to address the asymmetric deforma-
These four geometries were selected based on the premise that tion. Poisson’s ratio was calculated from each line according to
combining positive and negative Poisson’s ratio units can result in the
εx (Li − L0 )/L0 Li − L0
entire assembly exhibiting a zero Poisson’s ratio [30,31]. Additionally, νi = − = − = , i ∈ [1, 10]. (1)
εy δy /L0 δy
these non-periodic and asymmetric metamaterials hold great potential
for exhibiting multifunctionality that does not exist in periodic designs
Here, εx and εy are strains in the x and y directions, respectively. Li
[32]. The non-periodic lattice was generated by randomly assigning 64
represents the length of the ith line after deformation. L0 = 120 mm is
unit cells from this candidate pool to the 8 × 8 design space, as illus-
the initial length of each reference line. The average and standard de-
trated in Fig. 2. The initial training dataset (Strain) includes 2,000
viation of these 10 Poisson’s ratios were calculated as νavg and νstd,
randomly assembled metamaterial samples and their corresponding
respectively. Both of them were employed to characterize the asym-
Poisson’s ratios. Additionally, a validation dataset (Svalid) and a testing
metric deformation in the entire metamaterial. In this study, we consider
dataset (Stest) were created, each consisting of 500 new metamaterial
a sample exhibiting a zero Poisson’s ratio if it meets the criteria |νavg| <
samples. These datasets were generated using the same procedure. Svalid
0.05 and 0 ≤ νstd < 0.02. It can be found from Fig. 4 that only 6 designs
was used for monitoring model performance and preventing overfitting
out of a pool of 2,000 randomly generated training samples met the
during model training as discussed in Section 2.2. Stest was used to
design criteria. This type of training set cannot be handled by traditional
evaluate the model accuracy, as mentioned in Section 3.
ML generative models. An OOD target-oriented inverse design model, as
To transform the lattice metamaterial into suitable categorical data
detailed in Section 2.2.1, is introduced to address this challenge.
for the subsequent machine learning model, a four-channel one-hot
encoding approach was employed to represent the pattern of unit cell
candidates. Each channel was represented by an 8 × 8 matrix, corre- 2.2. Model description
sponding to the lattice assembly with 8 × 8 unit cells. Within each
channel, the presence or absence of the corresponding unit cell was 2.2.1. An inverse OOD target-oriented cVAE model for metamaterial design
denoted by “1” or “0” as illustrated in Fig. 2. For example, the meta- As illustrated in Fig. 5, we propose an OOD target-oriented cVAE
material sample in Fig. 2 had a reentrant unit cell located in the upper- model, a novel architecture with a custom training algorithm, for
left corner. Therefore, “1” was assigned to the corresponding location in generating metamaterial designs whose desired properties are rare (i.e.,
the reentrant channel, while “0” was assigned to the same location in the out-of-distribution) in the training dataset. Similar to a traditional cVAE
remaining three channels. This one-hot encoding scheme was applied model [34], our OOD target-oriented cVAE model comprises an encoder
consistently across the entire lattice. This approach ensures the con- hϕ and a decoder fθ, where ϕ and θ represent the trainable parameters of
nectivity of each unit cell, thus mitigating the risk of generating the encoder and the decoder, respectively. In our experiments, we used
unmanufacturable metamaterials. Moreover, the lattice dimension can an encoder with three convolutional layers and a decoder with five
be adjusted to accommodate various requirements and design specifi- transposed convolutional layers followed by another four convolutional
cations. The unit cell pool can be further enriched to create new designs. layers. The encoder hϕ transforms a metamaterial design x and its
The Poisson’s ratio of four periodic lattices and each metamaterial properties νavg and νstd into a variational posterior qϕ(z|x, νavg, νstd). This
was determined using a quasi-static uniaxial compression finite element posterior is defined as a Gaussian distribution with a diagonal covari-
simulation conducted with Abaqus [33]. Each unit cell was discretized ance matrix, characterized by its mean and the log-variances. These
into a grid of 30 × 30 pixel 2D quadratic elements to balance structural mean and log-variance values are the outputs of the encoder hϕ(x, νavg,
accuracy and computational efficiency. As depicted in Fig. 3(a), the νstd). Using the reparameterization trick [35], for a given metamaterial
metamaterial sample was positioned between two rigid plates in a design x and its properties νavg and νstd, we can sample a latent encoding
“surface-to-surface” contact configuration. The bottom plate was held z from the variational posterior qϕ(z|x, νavg, νstd), by first sampling an ∊
fixed while the top one was subjected to a displacement of δy = 6 mm. from a standard Gaussian distribution and then computing z = µ + ∊σ.
The density, Young’s modulus, Poisson’s ratio, and yield stress of the The decoder fθ transforms a latent encoding z along with the properties
bulk material were 103 kg/m3, 1850 MPa, 0.33, and 150 MPa, respec- νavg, νstd into a distribution pθ(x|z, νavg, νstd) parameterized by the output
tively. The bulk material always remained elastic during the simulation. of the decoder fθ(z, νavg, νstd). A metamaterial design x’ can then be
To verify the accuracy of the simulation, a 3D printed, randomly sampled from pθ(x|z, νavg, νstd).
selected metamaterial sample shown in Fig. 2 was mechanically tested. Metamaterial designs generated by traditional cVAE models often
In the setup, the 2D metamaterial was extruded by 3 mm to induce a utilize a pixelated image format to represent a metamaterial [36].
plane-strain state. The sample was fabricated using stereolithographic However, this format fails to capture the categorical information
additive manufacturing with a Phrozen 8K mini 3D printer and the inherent in assembled unit cells in our problem and is not aligned with
Phrozen Aqua Gray 8K resin. The printing resolutions along the x, y, and the metamaterial representations for our application, discussed in Sec-
z axes were 22 μm, 22 μm, and 50 μm, respectively. The exposure times tion 2.1, see Fig. 2. To be more precise, we parameterize pθ(x|z, νavg,
for the normal and bottom layers were set to 2.5 s and 30 s, respectively. νstd) by first assuming that the 8 × 8 unit cells are independent, such that
All other printer settings were kept at their default values. Following pθ(x|z, νavg, νstd) can be factorized as pθ(x|z, νavg, νstd) = Пi,j pθ(xi,j |z,
printing, the printed samples were cleaned with an ultrasonic cleaner for νavg, νstd), where xi,j is a one-hot vector for the (i, j)-th unit cell, denoting
5 minutes and UV-cured (Creality 3D UW-01) for 30 minutes. During the whether the (i, j)-th unit cell contains a square, honeycomb, reentrant, or
compression process, the 3D-printed metamaterial sample was triangle unit (We use xi,j(k) to denote the value of xi,j for the k-th type of

3
X. Liu et al. Materials & Design 244 (2024) 113224

unit — xi,j(k) = 1 or 0 denotes whether the k-th type of unit is used or not is precisely the cross-entropy loss, which maximizes the (log− ) likeli-
at the (i, j)-th unit cell). Instead of using a Gaussian (or a Bernoulli) hood of generating (i.e., reconstructing) the same type of unit at the (i,
distribution to parameterize pθ(xi,j |z, νavg, νstd), as done in the prior j)-th cell as that indicated by the training input xi,j at the same cell.
work [34], we used a categorical distribution to parameterize pθ(xi,j |z, Hence, we can interpret the first loss term as a reconstruction loss, which
νavg, νstd), where the distribution parameters were defined by the is implemented as a cross-entropy loss summed over all (i, j)-th unit cells.
decoder output fθ(z, νavg, νstd) at the (i, j)-th unit cell (denoted fθ(z, νavg, This again differs from the prior work, where a reconstruction loss is
νstd)i,j). The distribution pθ specifies the probabilities, for each of the 8 × generally implemented using a mean squared loss or a logistic (binary
8 unit cells, of having a square, honeycomb, reentrant, and triangle unit cross entropy) loss [38].
at that cell. In other words, the output of our decoder is an 8 × 8 The second loss term is the Kullback-Leibler (KL) divergence be-
probability matrix for each unit cell channel. We denote the k-th channel tween the variational posterior qϕ(z|x, νavg, νstd) and the prior distri-
of the decoder output as fθ(k)(z, νavg, νstd), and its (i, j)-th entry as fθ(k)(z, bution p(z). This term can be viewed as a regularization of the
νavg, νstd)i,j. The latter is the probability that the k-th type of unit is used reconstruction loss. Since both the variational posterior qϕ(z|x, νavg, νstd)
∑
at the (i, j)-th cell, and we have k fθ(k)(z, νavg, νstd)i,j = 1. In order to and the prior distribution p(z) over the latent encodings are Gaussian
∑
interpret fθ (z, νavg, νstd)i,j as a probability and to ensure k fθ(k)(z, νavg,
( k)
distributions, this loss term can be computed analytically:
νstd)i,j = 1, a softmax normalization layer was included to convert the
unnormalized decoder output to a probability vector at each unit cell. ( ( ) ) 1∑ J ( )
DKL qϕ z|x, vavg , vstd ‖p(z) = − 1 + logσ2j − μ2j − σ2j , (5)
To generate a metamaterial design x’ from pθ(x|z, νavg, νstd), we 2 j=1
sample one of the four units, for each of the 8 × 8 unit cells, according to
the probabilities fθ(z, νavg, νstd) produced by the decoder. This modifi- where µj and σj2 are the j-th component of µ and σ 2. Both were deter-
cation effectively prevents the generation of unmanufacturable meta- mined from the output of the encoder µ, log σ2 = hϕ(x, νavg, νstd).
materials, such as those with non-connected features like hanging nodes, Following β-VAE [37] and Eq. (2), the KL-divergence term was weighed
which cannot be produced using conventional manufacturing tech- with β = 0.001 × 64, where 64 is the total unit cell number; 0.001 is
niques. By ensuring manufacturability, our approach eliminates the selected during the parameter tuning process, as it yielded the best
need for additional algorithms to correct these issues. Moreover, this performance on the validation Svalid.
modification adds a more stochastic nature to the model for better In our experiments, we found that training a cVAE with a standard
exploration of the design space. stochastic gradient descent, where training instances are sampled uni-
Our model was trained by optimizing the β-VAE objective [37], formly and randomly from the training dataset, failed to generate met-
defined for minimization on a single training input (x, νavg, νstd) by: amaterial designs with the near-zero νavg and νstd. We believe that this is

( [ ( )] ( ( ) )
L ϕ, θ; x, vavg , vstd = − Eqϕ (z|x,vavg ,vstd ) logpθ x|z, vavg , vstd + βDKL qϕ z|x, vavg , vstd ‖p(z) , (2)

where p(z) is the prior distribution over the latent encodings z as a because our training dataset contained very few instances (0.3 %) that
standard Gaussian distribution. The first term on the right in Eq. (2) is had near-zero νavg and νstd – the vast majority of the training instances
the negative log-likelihood of generating the original design x from the do not fulfill the desired properties. To rectify this problem, an OOD
latent encoding z. z is sampled from the variational posterior qϕ(z|x, target-oriented sampler was used to sample training instances before
νavg, νstd) defined by the encoder output hϕ(x, νavg, νstd), averaged over passing them to the cVAE for training. The OOD target-oriented sampler
all possible latent encodings. This term can be approximated by taking a selects a training instance x with the properties of νxavg, νxstd, based on
Monte Carlo sample from the variational posterior qϕ(z|x, νavg, νstd). how close its properties are to the near-zero νavg and νstd (represented by
This approach is equivalent to passing the input x , νavg, νstd into the ν*avg and ν*std). The closer a training instance x’s properties of νxavg, νxstd
encoder to obtain a latent encoding z1 and then compute − log pθ(x|z, are to the near-zero νavg and νstd, the larger the probability of the
νavg, νstd) for that latent encoding z. Note that we have instance x being selected by the OOD target-oriented sampler for

( ) ∏ ( ) ∑ ( )
− logpθ x|z, vavg , vstd = − log pθ xi,j |z,vavg , vstd = − logpθ xi,j |z,vavg , vstd (3)
i,j i.j

training will be. Mathematically, the probability of the training instance

where xi,j is the one-hot vector for the (i, j)-th unit cell from the training x is proportional to:
input x. Since pθ(xi,j |z, νavg, νstd) is a categorical distribution, ⎛ ⎡( )2 ( ⎤⎞
)2
( ) 1 v x
avg − v *
avg v x
− v *
( ) 4
∑ [ ] ( ) f vxavg , vxstd = exp⎝ − ⎣ + std std ⎦ ⎠, (6)
− logpθ xi,j |z,vavg , vstd = − 핀 xi,j = 1 logf (k)
(k)
z, vavg , vstd i,j , (4) 2 σ vavg σ vstd
θ
k=1
For a given training dataset Strain, this probability is computed by:
( )
( ) f vxavg , vxstd
Pr x is selected |vxavg , vxstd = ∑ ( ). (7)
x’ ∈Strain f vavg , vstd
1 x’ x’
Recall that z is stochastic because it is computed as z = μ +∊σ where μ and
σ come from the encoder outputs hϕ (x, νavg, νstd) and ∊ is sampled from stan-
dard Gaussian. The σνavg and σνstd in f(νxavg, νxstd) are pre-defined hyperparameters

4
X. Liu et al. Materials & Design 244 (2024) 113224

Fig. 2. A metamaterial sample with its categorical data representation.

that control how close νxavg and ν*avg, and νxstd and ν*std are needed, new metamaterial designs. This high throughput tool significantly
respectively. This affects the probability that the instance x will be avoids the high cost and time associated with traditional property
selected by the sampler. As indicated in Eq. (6), with smaller values of evaluation methods. In this approach, the CNN architecture consists of
σ νavg and σνstd, training samples whose νxavg, νxstd are close to the desired multiple convolutional layers, each followed by batch normalization
values ν*avg, ν*std are more likely to be selected by the sampler. and the Rectified Linear Unit (ReLU) activation functions to form the
Here, the Effective Ratio (ER) is defined to control the learning focus convolutional group. Three sets of convolution groups and pooling
and quantify the difference in sampling probability between designs that layers are connected to reduce both the spatial dimension and output
meet the desired criteria and those that do not. Specifically, ER is channel by half. Two fully connected layers are then used to integrate
computed as: the extracted features with the predicted νavg and νstd. Detailed in this
∑ ( ) ∑ ( ) study, the CNN was implemented using PyTorch [40]. Adam Optimizer,
x∈S+ Pr x is selected |vxavg , vxstd x∈S+ f vxavg , vxstd based on stochastic gradient descent (SGD), was employed to minimize
ER = ∑ train
( )=∑ train
( ), (8)
Pr x’ is selected |vxavg
’
, vxstd
’
f vxavg
’
, vxstd
’ the mean square loss function which indicates the average of all the
x’ ∈S− x’ ∈S−
train train
squared differences between the true value and the predicted value. To
prevent overfitting, the early stopping technique was implemented to
where S+ train is the set of training samples that satisfy the zero Poisson’s terminate the CNN training when the validation loss reached its mini-
ratio, while S−train is the set of training samples that do not satisfy the zero mum. The performance of the forward CNN model was evaluated and is
Poisson’s ratio. In our case, when both acceptable and unacceptable presented in Section 3.1.
samples have an equal probability of being selected for training, ER = 1. As summarized in Fig. 7, our active learning strategy involves three
If the training needs to focus more on acceptable samples, then an ER > 1 steps. Step 1: Generate 200 samples using the OOD target-oriented cVAE
is required. model. Step 2: Calculate the corresponding Poisson’s ratios of the 200
Unlike other cVAE models, Our OOD target-oriented cVAE model samples using the forward CNN model. Here, the well-trained CNN
provides the flexibility to control ER and make “rare” samples “unrare” model was used to screen out unacceptable metamaterial designs from
during the training process, as shown in Fig. 6. As a result, the repre- the cVAE generation in Step 1. Step 3: Reevaluate the Poisson’s ratios of
sentation of acceptable samples is less constrained by their initial pro- the remaining samples using finite element simulation, as described in
portion in the training set. A more detailed discussion of how ER affects Section 2.1. The remaining acceptable samples are added to the training
the performance of our model will be further discussed in Section 3.2. set. This completes the first iteration of active learning. This cycle re-
In our application, the desired properties are ν*avg = 0 and ν*std = 0. peats until the model performance shows diminishing improvements.
If a training instance x’s properties, νxavg, νxstd, are close to the desired Incorporating the forward CNN model into the active learning process
properties ν*avg = 0 and ν*std = 0, f(νxavg, νxstd), and consequently, Pr(x can effectively recalibrate the machine learning focus during the
is selected|νxavg, νxstd) would have a large value. Such a training instance training process. A detailed discussion on how active learning enhances
will therefore be more likely to be selected by the OOD target-oriented model efficiency is provided in Section 3.3.
sampler for training. As shown by our experiments (Section 3), our
approach can effectively mitigate issues related to datasets with sparse 3. Results and discussion
target data during the training process.
3.1. Performance of the forward CNN model in predicting Poisson’s ratio
2.2.2. Active learning strategy
To further enhance the performance of the OOD target-oriented Fig. 8(a) depicts the mean square error loss as a function of epochs
cVAE model, an active learning strategy [39] was employed, as shown for Strain and Svalid during the training process. Upon reaching 100
in Fig. 7. In each iteration, the training dataset is continually enriched epochs, the overall training and validation loss decreased to 0.004. In
with new metamaterial samples that meet the design criteria. Since order to prevent overfitting, we consider the forward CNN model to be
calculating Poisson’s ratio for newly generated samples using finite- well-trained around 150 epochs. The performance of the well-trained
element simulations is highly inefficient (Section 2.1), we trained a CNN model was evaluated by predicting the νavg and νstd from
forward CNN model, which can rapidly predict νavg, νstd based on any

5
X. Liu et al. Materials & Design 244 (2024) 113224

Fig. 3. Comparison of the (a) stress-strain curve and (b) deformation history of a metamaterial sample from finite element simulation and in-situ compres-
sion experiment.

metamaterial samples in the test dataset Stest. As depicted in Fig. 8(b), only time-consuming but also dependent on resource availability. This
the dashed line corresponds to an ideal scenario where the νavg values raises two primary questions: (1) How much training data (N) is suffi-
predicted from the well-trained CNN model align with those from the cient for the machine learning model to provide reliable material design
corresponding training data. Since most of the test predictions are very feedback? (2) Given the constraints of limited training data, how can we
close to the ground truth values, as shown in Fig. 8(b), we can use the enhance the machine learning efficiency by optimizing the model? In
trained CNN as a surrogate model for predicting Poisson’s ratio. Quan- the following discussions, the acceptable rate, defined as the ratio of
titatively, the Root Mean Square Error (RMSE) of the CNN model is less generated designs meeting the criteria to the total number of generated
than 0.02, and its R2 value on the test dataset is close to 0.82 for νavg and designs, serves as a metric to evaluate the inverse design efficiency. The
0.64 for νstd. While better performance could potentially be achieved by effective ratio (ER), as introduced in Section 2.2.1, quantifies the
further optimizing the architecture and hyperparameters of the CNN training focus between samples that meet the design criteria and those
model, it does not impact the accuracy of the inverse design, as the that do not. When both acceptable and unacceptable samples have an
forward CNN model is solely used for the initial screening of unqualified equal probability of being selected for training, the effective ratio equals
samples. Samples that pass this screening must undergo verification one. A ratio greater than one indicates that training is more focused on
through finite element simulation. This step can further screen out the acceptable samples, while a ratio less than one indicates a stronger focus
false positives. on unacceptable samples.
In this study, we considered five sets of randomly generated training
data with numbers N = 300, 500, 1000, 2000, and 4000, respectively. It
3.2. Effect of training data size, effective ratio, and initial ratio on inverse can be observed from Fig. 9(a) that when N = 300, the acceptable rate is
design efficiency zero. This is because no sample in the training data meets the design
criteria. Based on our random data generation, only 6 out of 2,000
Acquiring training data through experiments or simulations is not

6
X. Liu et al. Materials & Design 244 (2024) 113224

Fig. 4. Distribution of νavg and νstd within the training dataset Strain.

samples can satisfy the design criteria. To ensure at least one acceptable performance, we considered an initial training set with N = 1,000 at an
sample is included during training, N should be greater than 2000/6 ≈ ER ≈ 1 for the following study.
333. Below this threshold, ER is always zero and cannot influence the As illustrated in Fig. 7, active learning is implemented by adding
learning process. However, beyond this threshold, ER starts to play its additional samples to the initial training set through the following steps.
role. As shown in Fig. 9(a), when ER < 10− 1, the acceptable rate remains Firstly, 200 new training samples were generated using the OOD target-
at 0 for the remaining four cases. This is because the probability of oriented cVAE model. Then, the Poisson’s ratios of the 200 new gener-
selecting acceptable samples is still less than one-tenth of that of unac- ations are first verified by the forward CNN model and then validated
ceptable ones. Our model cannot effectively learn from the acceptable through finite element simulation. Next, the n1 samples that passed the
samples if they are still severely underrepresented during the training verification were added to the initial training set, increasing the total
process. When ER increases from 10− 1 to 1, there is a significant number of training samples to N = 1000 + n1. This process repeats for
∑M
improvement in the acceptable rate due to the enhanced representation subsequent iterations with N = 1000 + i=1 ni , where M is the total
of acceptable samples. However, a further increase in ER does not number of active learning iterations.
necessarily lead to continued improvements in the acceptable rate. The As shown in Fig. 10(a), the acceptable rate significantly increased
analysis indicates that even if the training process exclusively focuses on from 17 % to 32 % after the first iteration. This corresponds to a tenfold
acceptable samples (when ER approaches infinity), it makes little dif- increase in IR, thereby enhancing the representation of acceptable de-
ference compared to equally selecting both acceptable and unacceptable signs during the training process. However, it is worth noting that after
samples for training (ER = 1). Overrepresentation of acceptable samples the second iteration, the acceptable rate starts to saturate even if the IR
cannot further improve learning as the model already extracted enough keeps increasing. This indicates that increasing the IR beyond a certain
information when ER ≈ 1. level does not further improve the acceptable rate. This phenomenon is
It is worth noting that when N = 1,000, 2,000, and 4,000, the similar to the effect of ER, as ER > 1 has a negligible impact on the
acceptable rate tends to stabilize at a similar level, which is about twice acceptable rate, according to the discussion in Section 3.2. It can be
as high as the acceptable rate observed when N = 500. This means that concluded that even in an extreme scenario where the training set only
the training data size itself cannot ultimately determine the model’s consists of acceptable samples, or the training focus is exclusively on
performance. It is important to look at the initial data composition. acceptable samples, our model still cannot guarantee the generation of
Here, we define the Initial Ratio (IR) as the ratio between the number of an acceptable metamaterial.
acceptable samples and the number of unacceptable samples in the This limitation arises because generative models cannot achieve 100
initial training set. It can be seen from Fig. 9(a) that IR remains at 3 × % accuracy, regardless of how we shift the training focus or modify the
10− 3 when N = 1,000, 2,000, and 4,000, but decreases to 2 × 10− 3 when composition of the training data through active learning [41]. The
N = 500, and further drops to 0 when N = 300. It can be concluded that reason is twofold. Firstly, while we provide the model with a one-to-one
more data does not always lead to better training outcomes. Improving correlation between the metastructure and Poisson’s ratio, we expect
data quality, represented by IR in this case, is more critical. As shown in the model to generate multiple metastructure designs for a given Pois-
Fig. 9(b), increasing the IR from 1 × 10− 3 to 3 × 10− 3 results in son’s ratio. This requires the model to incorporate some degree of
approximately a threefold enhancement of the acceptable rate when N randomness, as illustrated in Fig. 5, which is reflected through the
= 1,000. However, it’s important to note that the maximum attainable Gaussian distribution in Eq. (2), the Monte Carlo sampling in Eq. (3),
IR for a randomly generated training set is 3 × 10− 3. Further improve- and the categorical distribution in Eq. (4). Secondly, this randomness
ments require active learning, as discussed in Section 3.3 below. inevitably introduces inaccuracies [42], as the model may generate
metamaterials that are not acceptable. While not perfect, the active
learning strategy can significantly enhance model efficiency by reducing
3.3. Effect of the active learning strategy on inverse design efficiency
the time required to identify acceptable metamaterial designs. As
depicted in Fig. 10(b), our OOD target-oriented cVAE model requires
Based on the discussion in Section 3.2, increasing the training data
only 2 hours compared to approximately 4 hours with the traditional
size N beyond 1,000 yields negligible improvement in the stabilized
cVAE model and a substantial 160 hours when selecting from random
acceptable rate. To balance efficiency in data acquisition and model

7
X. Liu et al. Materials & Design 244 (2024) 113224

Fig. 5. Architecture of the OOD target-oriented cVAE model.

Fig. 6. Schematic illustration of using OOD target-oriented sampler to adjust the Effective Ratio (ER).

8
X. Liu et al. Materials & Design 244 (2024) 113224

Fig. 7. Scheme of the active learning strategy.

Fig. 8. (a) Loss-epoch curve of the CNN model; (b) Distribution of the predicted νavg values from the forward CNN model using the test set Stest.

Fig. 9. Effect of (a) training data size N, effective ratio ER, and (b) initial ratio IR on the acceptable rate.

generation methods. Additionally, as shown in Fig. 11, our model can design models, our approach can handle a training dataset where the
generate 152 acceptable designs after 5 active learning iterations, majority of samples do not meet the design criteria (few-shot). This
whereas the random generation approach yields only 1 design using the functionality was achieved through an out-of-distribution (OOD) target-
same training set with N = 400. oriented Conditional Variational Autoencoder (cVAE) model enhanced
It can be concluded that implementing active learning into the in- by an active learning strategy. When the target design is extremely out-
verse OOD target-oriented cVAE model is an effective solution for of-distribution in the training dataset, simply increasing the number of
overcoming the challenge of data scarcity in target designs, although it training data does not always improve the efficiency of the machine
has a limit in achieving perfect accuracy due to inherent randomness in learning model. Instead, an effective sampler and/or an active learning
generative models. In addition to Poisson’s ratio, this model can be strategy should be employed to ensure a balanced selection of targeted
applied to discover novel metamaterial designs with other properties. and untargeted designs during the training process. Fine-tuning this shift
Our future work will focus on refining and expanding the capacity of this in learning emphasis can be achieved by adjusting the effective ratio
model to fracture, sensing, and energy harvesting applications [43–45]. (ER) through the OOD target-oriented sampler and a few iterations of
active learning. Compared to the random trial-and-error approach, our
4. Conclusion OOD target-oriented cVAE model with an active learning strategy ach-
ieved a remarkable 87.5 % reduction in time and a significant 130-fold
In this study, a generative framework was developed for designing increase in design efficiency. This model provides crucial design
metamaterials with zero Poisson’s ratio using few-shot learning. Meta- guidelines for innovative non-periodic metamaterials without the need
material samples were created by randomly assigning four-unit cell for repetitive analysis or human intervention. The insights gained from
candidates within an 8 × 8 design space. Compared to other inverse this study are applicable to scaled-up design layouts, such as 16 × 16

9
X. Liu et al. Materials & Design 244 (2024) 113224

Fig. 10. (a) Effect of active learning iterations on acceptable rate (initially, N = 1000); (b) Comparison of time required to generate 50 acceptable metamaterials
using different approaches.

Fig. 11. Comparison of the number of acceptable designs using (a) random generation and (b) the inverse OOD target-oriented cVAE model after 5 iterations. Dataset
is N = 400 for both cases.

metamaterials, through domain decomposition. In our future work, we that could have appeared to influence the work reported in this paper.
plan to leverage physics-informed machine learning models [15] to
uncover the underlying local patterns. This approach will allow us to Data availability
accumulate and transfer our knowledge base effectively.
The datasets and code used during the current study are available
CRediT authorship contribution statement from the corresponding author upon reasonable request.

Xiangbei Liu: Writing – original draft, Visualization, Validation, Acknowledgments

Investigation, Formal analysis, Data curation, Methodology, Conceptu-
alization, Software. Huan Zhao: Writing – review & editing, Visuali- The authors would like to acknowledge the support from the Na-
zation, Formal analysis. Ya Tang: Writing – review & editing, tional Aeronautics & Space Administration (NASA) under Award
Visualization, Data curation. Chaofan Chen: Writing – review & edit- #80NSSC22M0175, the National Science Foundation under Award
ing, Supervision, Methodology, Conceptualization. Yifeng Zhu: Writing #OIA-2218063, as well as the start-up funds from Thayer School of
– review & editing, Supervision, Resources, Funding acquisition, Engineering at Dartmouth College.
Conceptualization. Bo Song: Writing – review & editing, Resources. Yan Sandia National Laboratories is a multi-mission laboratory managed
Li: Writing – review & editing, Writing – original draft, Supervision, and operated by National Technology & Engineering Solutions of San-
Software, Resources, Project administration, Funding acquisition, dia, LLC (NTESS), a wholly owned subsidiary of Honeywell International
Conceptualization. Inc., for the U.S. Department of Energy’s National Nuclear Security
Administration (DOE/NNSA) under contract DE-NA0003525. This
written work is authored by an employee of NTESS. The employee, not
Declaration of competing interest NTESS, owns the right, title and interest in and to the written work and is
responsible for its contents. Any subjective views or opinions that might
The authors declare the following financial interests/personal re- be expressed in the written work do not necessarily represent the views
lationships which may be considered as potential competing interests: of the U.S. Government. The publisher acknowledges that the U.S.
Xiangbei Liu, Huan Zhao reports financial support was provided by Government retains a non-exclusive, paid-up, irrevocable, world-wide
NASA. Xiangbei Liu reports financial support was provided by National license to publish or reproduce the published form of this written
Science Foundation. Bo Song reports financial support was provided by work or allow others to do so, for U.S. Government purposes. The DOE
US Department of Energy. Yan Li reports financial support was provided will provide public access to results of federally sponsored research in
by NASA. Yan Li reports financial support was provided by National accordance with the DOE Public Access Plan.
Science Foundation. If there are other authors, they declare that they
have no known competing financial interests or personal relationships

10
X. Liu et al. Materials & Design 244 (2024) 113224

Appendix A. Supplementary data [21] I. Tanriover, D. Lee, W. Chen, K. Aydin, Deep generative modeling and inverse
design of manufacturable free-form dielectric metasurfaces, ACS Photon. (2022).
[22] A. Challapalli, D. Patel, G. Li, Inverse machine learning framework for optimizing
Supplementary data to this article can be found online at https://doi. lightweight metamaterials, Mater. Des. 208 (2021).
org/10.1016/j.matdes.2024.113224. [23] X. Zheng, T.-T. Chen, X. Guo, S. Samitsu, I. Watanabe, Controllable inverse design
of auxetic metamaterials using deep learning, Mater. Des. 211 (2021).
[24] M.L. Ruh, X. Liu, R. Yu, J.T. Hwang, Airfoil shape parameterization using
References reconstruction-error-minimizing generative adversarial networks. AIAA AVIATION
2023 Forum, 2023.
[1] A.J. Lew, K. Jin, M.J. Buehler, Designing architected materials for mechanical [25] M. Abdollahzadeh, T. Malekzadeh, C.T. Teo, K. Chandrasegaran, G. Liu, N.-M.
compression via simulation, deep learning, and experimentation, NPJ Comput. Cheung, A survey on generative modeling with limited data, few shots, and zero
Mater. 9 (1) (2023). shot, 2023. arXiv preprint arXiv:2307.14397.
[2] J. Fan, L. Zhang, S. Wei, Z. Zhang, S.-K. Choi, B. Song, Y. Shi, A review of additive [26] T. Karras, M. Aittala, J. Hellsten, S. Laine, J. Lehtinen, T. Aila, Training generative
manufacturing of metamaterials and developing trends, Mater. Today 50 (2021) adversarial networks with limited data, Adv. Neural Inf. Proces. Syst. 33 (2020)
303–328. 12104–12114.
[3] R.M. Gorguluarslan, U.N. Gandhi, R. Mandapati, S.-K. Choi, Design and fabrication [27] Y. Wang, Q. Yao, J.T. Kwok, L.M. Ni, Generalizing from a few examples, ACM
of periodic lattice-based cellular structures, Comput.-Aided Des. Applic. 13 (1) Comput. Surv. 53 (3) (2020) 1–34.
(2015) 50–62. [28] H. Yang, B. Wang, L. Ma, Designing hierarchical metamaterials by topology
[4] X. Liu, J.J. Jeon, A.G. Tiplea, Y. Li, B. Song, Design of low density architectured analysis with tailored Poisson’s ratio and Young’s modulus, Compos. Struct. 214
metamaterials with high compressive and torsional stiffness. Proc. ASME (2019) 359–378.
International Mechanical Engineering Congress and Exposition, American Society [29] B. Ling, K. Wei, Z. Wang, X. Yang, Z. Qu, D. Fang, Experimentally program large
of Mechanical Engineers, 2023 p. V004T004A009. magnitude of Poisson’s ratio in additively manufactured mechanical
[5] N. Gaspar, C.W. Smith, K.E. Evans, Effect of heterogeneity on the elastic properties metamaterials, Int. J. Mech. Sci. 173 (2020).
of auxetic materials, J. Appl. Phys. 94 (9) (2003) 6143–6149. [30] C. Minfang, Z. Yifeng, L. Rong, W. Shiwen, I.A. Evrard, Equivalent-oriented model
[6] J.N. Grima, L. Oliveri, D. Attard, B. Ellul, R. Gatt, G. Cicala, G. Recca, Hexagonal for sandwich panels with ZPR accordion honeycomb, Int. J. Mech. Sci. 268 (2024).
honeycombs with zero Poisson’s ratios and enhanced stiffness, Adv. Eng. Mater. 12 [31] K.R. Olympio, F. Gandhi, Zero Poisson’s ratio cellular honeycombs for flex skins
(9) (2010) 855–862. undergoing one-dimensional morphing, J. Intell. Mater. Syst. Struct. 21 (17)
[7] K.A. Khan, M.A. Khan, 3-3 piezoelectric metamaterial with negative and zero (2009) 1737–1753.
Poisson’s ratio for hydrophones applications, Mater. Res. Bull. 112 (2019) [32] H. Zhao, X. Liu, Y. Li, Architecture design of high-performance piezoelectric energy
194–204. harvester with 3D metastructure substrate, Adv. Theory Simul. p. 2301214.
[8] Y. Chen, M.H. Fu, Mechanical properties of a novel zero Poisson’s ratio [33] M. Smith, ABAQUS/CAE, Dassault Syst{\‘e}mes Simulia Corp, United States, 2022.
honeycomb, Adv. Eng. Mater. 20 (2) (2017). [34] K. Sohn, H. Lee, X. Yan, Learning structured output representation using deep
[9] D. Attard, J.N. Grima, Modelling of hexagonal honeycombs exhibiting zero conditional generative models, in: Advances in Neural Information Processing
Poisson’s ratio, Physica Status Solidi (b) 248 (1) (2010) 52–59. Systems, 2015, p. 28.
[10] A. Farrokhabadi, M.M. Ashrafian, M. Fotouhi, Design and characterization of an [35] D.P. Kingma, S. Mohamed, D. Jimenez Rezende, M. Welling, Semi-supervised
orthotropic accordion cellular honeycomb as one-dimensional morphing structures learning with deep generative models, Adv. Neural Inf. Process. Syst. 27 (2014).
with enhanced properties, J. Sandw. Struct. Mater. 24 (3) (2022) 1726–1745. [36] L. Wang, Y.-C. Chan, F. Ahmed, Z. Liu, P. Zhu, W. Chen, Deep generative modeling
[11] J.K. Wilt, C. Yang, G.X. Gu, Accelerating auxetic metamaterial design with deep for mechanistic-based learning and design of metamaterial systems, in: Computer
learning, Adv. Eng. Mater. 22 (5) (2020). Methods in Applied Mechanics and Engineering, 2020, p. 372.
[12] M. Fortes, M.T. Nogueira, The poison effect in cork, Mater. Sci. Eng. A 122 (2) [37] I. Higgins, L. Matthey, A. Pal, C.P. Burgess, X. Glorot, M.M. Botvinick, S. Mohamed,
(1989) 227–232. A. Lerchner, beta-vae: learning basic visual concepts with a constrained variational
[13] M. Hanifpour, C.F. Petersen, M.J. Alava, S. Zapperi, Mechanics of disordered framework, ICLR (Poster) 3 (2017).
auxetic metamaterials, Eur. Phys. J. B 91 (11) (2018). [38] J. Bao, D. Chen, F. Wen, H. Li, G. Hua, CVAE-GAN: fine-grained image generation
[14] B. Henderson, J.P.M. Whitty, P. Myler, C. Chirwa, Crash performance of cellular through asymmetric training, in: Proc. Proceedings of the IEEE International
foams with reduced relative density part 2: rib deletion, Int. J. Crashworthiness 12 Conference on Computer Vision, 2017, pp. 2745–2754.
(6) (2007) 689–698. [39] B. Settles, “From theories to queries: active learning in practice,” active learning
[15] J.M. Podestá, C.G. Méndez, S. Toro, A.E. Huespe, J. Oliver, Material design of and experimental design workshop in conjunction with AISTATS 2010, in:
elastic structures using Voronoi cells, Int. J. Numer. Meth. Eng. 115 (3) (2018) G. Isabelle, C. Gavin, D. Gideon, L. Vincent, S. Alexander (Eds.), PMLR,
269–292. Proceedings of Machine Learning Research, 2011, pp. 1–18.
[16] T. Maconachie, M. Leary, P. Tran, J. Harris, Q. Liu, G. Lu, D. Ruan, O. Faruque, [40] S. Imambi, K.B. Prakash, G. Kanagachidambaresan, PyTorch, Programming with
M. Brandt, The effect of topology on the quasi-static and dynamic behaviour of TensorFlow: Solution for Edge Computing Applications, 2021, pp. 87–104.
SLM AlSi10Mg lattice structures, Int. J. Adv. Manuf. Technol. 118 (11–12) (2021) [41] J. Bose, R.P. Monti, A. Grover, Controllable generative modeling via causal
4085–4104. reasoning, Trans. Mach. Learn. Res. (2022).
[17] A.P. Garland, B.C. White, S.C. Jensen, B.L. Boyce, Pragmatic generative [42] S. Mishra, S. Flaxman, T. Berah, M. Pakkanen, H. Zhu, S. Bhatt, πVAE: encoding
optimization of novel structural lattice metamaterials with machine learning, stochastic process priors with variational autoencoders, arXiv preprint arXiv:
Mater. Des. 203 (2021). 2002.06873, 970, 2020.
[18] J. Tian, K. Tang, X. Chen, X. Wang, Machine learning-based prediction and inverse [43] H. Zhao, J.E. Huddy, W.J. Scheideler, Y. Li, Rational design of 3D-printed
design of 2D metamaterial structures with tunable deformation-dependent metastructure-based pressure sensors, Adv. Eng. Mater. (2023).
Poisson’s ratio, Nanoscale 14 (35) (2022) 12677–12691. [44] J.E. Huddy, A.P. Tiwari, H. Zhao, Y. Li, W.J. Scheideler, Graph theory design of 3D
[19] Q. Zeng, Z. Zhao, H. Lei, P. Wang, A deep learning approach for inverse design of printed conductive lattice electrodes, Adv. Mater. Technol. 8 (15) (2023).
gradient mechanical metamaterials, Int. J. Mech. Sci. 240 (2023). [45] H. Zhao, X. Liu, Y. Li, The role of fracture patterns on crack-based strain sensors,
[20] N.A. Alderete, N. Pathak, H.D. Espinosa, Machine learning assisted design of shape- J. Eng. Mater. Technol. 146 (1) (2023).
programmable 3D kirigami metamaterials, NPJ Comput. Mater. 8 (1) (2022).

11