Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

Materials Today Communications 37 (2023) 107407

Contents lists available at ScienceDirect

Materials Today Communications


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

Geometrical recognition of metallic foam microstructures using a deep


learning approach
William D. Romero a , Santiago Torres-Bermudez b , Brayan Valenzuela a , Cristian C. Viáfara c ,
Angel M. Meléndez b , Fabio Martínez a ,∗
a
Biomedical Imaging, Vision and Learning Laboratory (BIVL2 ab), Universidad Industrial de Santander (UIS), 680002, Colombia
b Grupo de Investigaciones en Minerales, Biohidrometalurgia y Ambiente (GIMBA), Universidad Industrial de Santander (UIS), 680002, Colombia
c Grupo de Investigaciones en corrosión (GIC), Universidad Industrial de Santander (UIS), 680002, Colombia

ARTICLE INFO ABSTRACT

Keywords: The geometric characteristics of nanostructured porous metallic foams (NMFs) impact their properties, enabling
Porous metals applications in electrocatalysis, energy conversion, and energy storage. The morphology of these three-
Electrochemical deposition dimensional (3D) structures is complex, in large part, because of their complicated formation mechanisms. This
Metal nanofoams
work presents a computational strategy for geometrically characterizing different NMFs designs. Three types
Auto-supervised architectures
of samples were synthesized using the dynamic hydrogen bubble template electrodeposition method. A set of
Deep representations
Embedded vectors
microscopic confocal images of NMFs was used as training input samples. Herein, a variational autoencoder
(VAE) was adjusted under a pretext task, allowing learning embedding descriptors to represent the geometry
of microscopic observations. To evaluate the feasibility of using a VAE to assist in the classification of NMFs,
a set of machine learning classifiers was implemented to discriminate among the groups of embedded vectors
according to their classes. Furthermore, the capacity of the VAE was validated regarding the capability to
separate the vectors according to their classes. Explainability mechanisms stand out geometrical features of
input images that had major support during the classification task.

1. Introduction possibilities for material development [6,7]. Integration of deep learn-


ing techniques with the study of metal nanofoams presents exciting
Metal nanofoams have emerged as a fascinating and promising prospects for accelerating their discovery, optimization, and tailored
class of materials in recent years, thanks to their unique porous struc- design.
tures and remarkable properties, which include high surface area, Computer vision techniques allow, among other things, quantifying
increased surface reactivity, and tunable morphology and composi- properties, avoiding observation bias, and reducing costs associated
tion [1]. These networks of metallic nanoparticles, combined with with the characterization of porous materials. For example, Chung et al.
interconnected porous structures, bring substantial advantages to elec- estimated the mechanical properties of foamed samples by analyzing
tron and mass transport processes. This unique set of attributes makes X-ray micro-computed tomography (micro-CT) images [8]. In these
metal nanofoams highly versatile materials with promising applications
sequences, descriptors for pore circularity and relative density were
in electrochemical applications, ranging from catalysis to sensors and
designed from binarized images, which were then validated with exper-
energy storage to sustainable technologies [1–3]. Dynamic Hydrogen
imental results. Liu et al. demonstrated the existence of a relationship
Bubble Templated (DHBT) electrodeposition is an innovative electro-
between the geometric parameters and the mechanical properties of
chemical method that has garnered interest due to its operation at
NMFs, following geometrical properties in binarized images taken from
atmospheric pressure and ambient temperature in a short period. By
leveraging the dynamic behavior of hydrogen bubbles at the electrode scanning electron microscope [9]. Other approaches have correlated
surface, DHBT offers precise control over the deposition process, en- microstructural properties with patterns in microscopic images using
abling the synthesis of complex and customizable structures [1,2,4,5]. deep representations [6]. Hiraide et al. proposed a methodology to
The convergence of artificial intelligence, particularly deep learning, predict Young’s module of a polymer from microstructural observa-
has revolutionized various scientific domains, unlocking unprecedented tions [7]. To achieve this, a generative adversarial network (GAN)

∗ Corresponding author.
E-mail address: famarcar@saber.uis.edu.co (F. Martínez).

https://doi.org/10.1016/j.mtcomm.2023.107407
Received 27 August 2023; Received in revised form 13 October 2023; Accepted 22 October 2023
Available online 24 October 2023
2352-4928/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
W.D. Romero et al. Materials Today Communications 37 (2023) 107407

Fig. 1. Schematic illustration of nanofoam synthesis by the dynamic hydrogen template method.
Source: Adapted from Plowman et al. [11].

Table 1 Each embedding represents a NMF class within a sample group that
Sample labels and chemical composition of the electrodeposition baths.
shares chemical compositions. Thus, the embedding representation and
Sample label Molar concentration (mol/L)
spatial attention maps allow support for geometrical characterization.
𝐶𝑢2+ 𝑁𝑖2+ The general pipeline of the proposed approach is illustrated in Fig. 3.
MF1 0.30 0.10
MF2 0.25 0.15
MF3 0.20 0.20 3.1. A variational autoencoder deep representation

Herein the main interest is to capture geometrical features from


learns material structure distribution from a latent space, while a observed NMFs without an explicit adjustment about explicit shape
convolutional net outputs mechanical properties. signatures, such as pore diameter, pore distribution, and perimeters,
Taking into account that the microstructure of NMFs is the most im- among others. Hence, an autoencoder architecture was adjusted, fol-
portant feature for high-performance electrochemical technologies be- lowing a pretext task for image reconstruction. Specifically, we take
cause it affects the efficiency of electrochemical reactions and advantage of a VAE to adjust latent vectors as a set of Normal dis-
convective-diffusive mass transport [3], this work introduces a deep tributions with [𝜇𝑖 , 𝜎𝑖 ] [12,13]. This architecture is trained with NFM
autoencoder representation that encodes the microstructure patterns of images using different chemical compositions and without class label
metallic copper–nickel nanofoams, as observed in microscopic images. information. We hypothesize that such embedding descriptors summa-
This representation was constructed using a variational autoencoder rize geometrical nanofoam’s features, expressed as a set of Gaussian
(VAE), and it is capable of describing hidden geometrical features de- distributions. Specifically, the VAE is composed of an encoder that takes
rived from embedding vectors learned during the image reconstruction input images 𝑥 and turns them into a lower-dimensional latent variable
task. The embedding vector, as a geometrical representation, enables 𝑧. Then, a decoder branch is charged to recover images 𝑥, ̂ following a
the classification of copper–nickel (Cu–Ni) nanofoams. binary-cross entropy reconstruction loss [14], expressed as in Eq. (1):

1 ∑∑
𝑀 𝑛
2. Dataset
𝑅𝐿𝑜𝑠𝑠 = − [𝑥 𝑙𝑜𝑔(𝑥̂ 𝑗,𝑖 ) + (1 − 𝑥𝑗,𝑖 )𝑙𝑜𝑔(1 − 𝑥̂ 𝑗,𝑖 )] (1)
𝑀 𝑖=1 𝑗=1 𝑗,𝑖
The NMFs were prepared following the dynamic hydrogen bubble
template method (DBTH) [5]. This method involves the use of a copper Particularly, the VAE architecture adjusts latent representation to a
substrate (cathode), which is immersed in an electrochemical cell filled standard Gaussian-probabilistic distribution  (𝑂, 𝐼) with mean 𝜇 = 0
with an acid aqueous solution containing the dissolved metal cations to and standard deviation 𝜎 equal to the identity matrix (𝐼) [13,15].
be deposited. When an electric current of 3.0 A∕cm2 is applied to the The VAE architecture is also adjusted concerning the Kullback–Leibler
cathode, hydrogen bubbles are generated, creating empty spaces on the divergence, described as:
cathode surface (porous space), and metal cations are transformed to

𝐾
their metal form on the substrate. This method is an electrochemical 𝐾𝐿𝑙𝑜𝑠𝑠 = −0.5 ∗ (1 + 𝑙𝑜𝑔(𝜎 2 ) − 𝜇 2 − 𝜎 2 ) (2)
process that allows the formation of NMFs [10], as illustrated in Fig. 1. 𝑖=1
Three types of NMF, hereafter called MF1, MF2, and MF3, were syn-
The minimization of this error implies that the values of the latent
thesized from three different aqueous solutions in 1.5 M H2 SO4 with
spaces variables tend to be similar to a Gaussian distribution, allowing
different ratios of CuSO4 ⋅ 5H2 O and NiSO4 ⋅ 6H2 O. Table 1 shows the
the model to build a continuous latent space that contains the spatial
composition of NMF types.
distribution of the visual features of the NMF images [16,17]. It should
These NMFs were deposited on a flat copper substrate with a
geometrical area of 1 cm2 . Each type of NMF was characterized using a be noted that the decoder should input a one-dimensional vector until
confocal stereoscopic microscope with a magnification of 700x. A total achieving a 𝑥 image reconstruction, making it easier to backpropagate
of 600 images of 1600 × 1200 pixels were recorded for the three NMFs the error. Then, the 2D embedding vectors (𝜇𝑖 , 𝜎𝑖 ) are reparametrized
classes, forming a balanced dataset, i.e., 200 samples for each class. as a linear combination that outputs a 𝑍 sampling vector [17,18],
Typical NMF samples for each composition are illustrated in Fig. 2. expressed as:

𝑧 = 𝜇 + 𝑙𝑜𝑔(𝜎) ∗ 𝜖 ∀ 𝜖 ∈  (0, 𝐼) (3)


3. Proposed approach
Hence, the VAEs are trained with a total error equation, which was
A Variational Autoencoder (VAE) encoding representation was introduced as the linear combination of both errors, as can be seen
trained to reconstruct the NMFs’ microscopic images. The adjusted VAE in Eq. (4).
generates embedding nanofoam descriptors, representing a latent repre-
sentation that codes multiple normal distributions of hidden features. 𝐸𝐿𝐵𝑂 = 𝐾𝐿𝐿𝑜𝑠𝑠 + 𝑅𝐿𝑜𝑠𝑠 (4)

2
W.D. Romero et al. Materials Today Communications 37 (2023) 107407

Fig. 2. Typical copper–nickel metallic foams microscopic images of each class in the dataset.

Fig. 3. Computational methodology for structural analysis of NMFs images. First, a VAE is trained. Next, the embedded vectors are projected to explore their conformation in a
low-dimensional latent space. In addition, a set of explainability maps supports the visual inspection task on the learned geometric features.

3.2. Embedding nanofoam descriptors among classes by quantifying the variance in the data [21]. For this
purpose, two different types of variances are taken into account: the
The latent space regularization obtained from the VAE allows the within-class variance (𝑆𝑤𝑖 ), and the between-class variance (𝑆𝑏𝑖 ), which
creation of a descriptor that encodes the geometric and structural prop- are measures of the dispersion of the data from the same and from
erties of the input NMF images. The set of learned normal distributions different classes, respectively. The 𝑊 transformation matrix of the LDA
then forms a latent space that can be explored for property recognition can be calculated as:
in materials, searching for elements that share similar mechanical
properties or measuring the degree of alloying between elements that 𝑆𝑊 𝑊 = 𝜆𝑆𝐵 𝑊 (5)
aid in synthesizing NMFs. In addition, the generation of new synthetic
where 𝜆 represents the eigenvalues of the transformation matrix 𝑊 .
embeddings taken from the latent space can simulate the conformation
and synthesis of new NMFs [19]. By selecting the 𝑘 most important eigenvalues, we can construct a
Once the visual descriptors are obtained from VAE as embedded matrix 𝑉𝑘 containing the eigenvectors associated with those selected
vectors that provide particular responses to input images, their sub- eigenvalues. From this new matrix, it is possible to project the original
sequent analysis can be performed through dimensionality reduction. points 𝑋 to a new lower-dimensional space 𝑌 , as shown in Eq. (6).
This dimensionality reduction allows, among others things, the recog-
𝑌 = 𝑋 ∗ 𝑊𝐾 (6)
nition of the data’s geometry the observation of possible patterns that
correlate with material properties, and can also serve as activation
maps to retrieve the most salient information from the learned vectors. 3.3. Explainable attention maps
Thus, we can effectively address the above problems using techniques
like PCA or LDA. Two dimensional reduction strategies were herein To support geometrical analysis from the embedding descriptors, in
implemented, described as follows this work, these learned distributions were backpropagated to the orig-
Unsupervised dimensional-reduction from PCA. This procedure is used inal image to determine principal regions associated with a particular
for dimensional reduction of information computing principal compo- descriptor. A modification to the Grad-CAM methodology was imple-
nents as linear combinations of the initial variables [20]. Each principal mented here to recover attention maps from embedding distributions
component, from an eigenvector decomposition, is a linear combination (see in Fig. 4).
of the original dimensions, where, the first component has the highest Particularly, for each element 𝑧𝑖 , the gradients are propagated with
variance of the points. respect to the features map 𝐴 ∈ R𝑁×𝐻×𝑊 , to obtain the activation map
𝑀 𝑖 [22].
Supervised dimensional-reduction from LDA. A Linear Discriminant Anal-
ysis (LDA) is a supervised method for dimensional reduction. This ∑
𝑛
𝑀 𝑖 = 𝑅𝑒𝐿𝑈 ( 𝛼𝑘 𝐴𝑘 ) (7)
technique is used to find new bases that present a better discrimination
𝑘=1

3
W.D. Romero et al. Materials Today Communications 37 (2023) 107407

Fig. 4. The application of Grad-CAM to a VAE is different from the application to common convolutional classification networks. In this case, since it does not output an activation
of a softmax layer, the Grad-CAM prediction is not based on class probability. Instead, what it seeks is to find the zones of interest learned by each dimension of the embedded
vector 𝑧.

𝛿𝑧
where the scalar 𝛼𝑘 is the global average 𝐺𝐴𝑃 ( 𝛿𝐴𝑖 ) of each element 𝑧𝑖 5. Evaluation and results
𝑘
with respect to the 𝑘th channel of the features map 𝐴 of the selected
convolutional layer calculated, as: For validation, to classify resultant embeddings, three conventional
machine learning architectures were used: KNN, Random Forest, and
1 ∑ ∑ 𝛿𝑧𝑖
ℎ 𝑤
𝛼𝑘 = ( ) (8) Decision Trees. As a computational baseline, a simple autoencoder was
𝑤𝐱ℎ 𝑝=1 𝑞=1 𝛿𝐴𝑝𝑞 implemented with the same layer configuration and hyperparameters
𝑘
as the VAE architecture. For this evaluation, we changed the embedded
where 𝐴𝑝𝑞𝑘
is the value of a pixel at position (𝑝, 𝑞). This process is re- vector size, using 512, 256 and 128 dimensions. With each of these
peated for each of the elements of the embedded vector [𝑧1 , 𝑧2 , … , 𝑧𝐷 ], models, the capability to classify between NMFs was evaluated using
which allows us to obtain an activation map for each element in the the mean 𝜇, variance 𝜎, and latent vector 𝑧 generated by the VAE.
embedded vector. In addition, it is also possible to obtain a total The ability to discriminate the embedded vectors from the conventional
saliency map from any linear combination between the individual architecture was also evaluated. Table 2 summarizes the achieved result
maps. Hence, we can obtain a single saliency map 𝑀 by averaging all for each embedding configuration and regarding each classifier. The
the individual maps. analysis of each of the components evidences the ability to geometri-
cally characterize NMFs based on the distribution of features within an
1 ∑ 𝑖
𝐷
𝑀= 𝑀 (9) embedded space.
𝐷 𝑖=1 The best classification results were obtained for the embedded
These saliency maps pay more attention to the regions or structures vector of the variance (𝜎). This discriminative character may be asso-
that are encoded in the embedded vector. In the application of NMFs ciated with the difference between learned distributions, based on the
analysis, it can be a relevant tool to determine the importance and dispersion of their data. That is, the distributions between classes can
be separated by their dispersion, but they share similar mean values.
contribution of each of the learned distributions, using the mean value
It can be observed an inverse relationship between the size of the
of each distribution as an approximation. For example, experts can
embedding the accuracy of the models to classify the NMFs. So, low
identify differential relationships in the geometry and structure of the
128-dimensional vectors are sufficient to represent the problem in the
NMFs, which can lead them to hypothesize new nanofoam syntheses
embedded space.
and investigate different chemical compositions.
Subsequently, the resulting embedded vectors (dimension 128) were
projected to a low-dimensional space from unsupervised PCA. Fig. 5
4. Experimental setup illustrates the representations of each type of embedded vector (𝜇, 𝜎, 𝑧),
and the geometric space in both the training (right column) and eval-
The encoder of the proposed architecture receives 128 × 128 images uation (left column) phases. As can be seen, the 𝜎 vector has a better
which are coded by four 2D convolutional layers. The first three layers geometrical representation, achieving a clearer separation between the
of the network have an associated MaxPooling2D operation and normal- three different compositions of the NMFs. Thus, geometric features can
ization of the activations (Batch Normalization). The number of filters be well encoded without label information about the composition. In
associated with each layer are 512, 256, 128, and 64, respectably. contrast, the projection of the 𝜇 and 𝑧 vectors does not prove to be
The last layer was projected to the embedded vectors representing the discriminative in either group, both training and evaluation.
distributions [𝜇, 𝜎] with 64 units for each parameter. Additionally, a Also, in a low-dimensional experiment, the embedded vectors were
linear combination of the last two dense layers was performed to obtain projected to a two-dimensional space, but following a supervised
a 𝑧 vector of 64 elements. scheme, through the LDA vector. The resulting embedded spaces for
each projection, both in training and evaluation, can be seen in Fig. 6.
The decoder receives 64 elements, which are progressively con-
In this case, the NFM type label has an direct impact on the dis-
volved (with filters of size 3 × 3) using 2D upsampling layers and a
crimination of the embedded space, resulting in geometrics with well-
batch normalization layer. The last layer was dedicated to recovering
separated features. The impact on minimization in this case has a
the original representation (128 × 128 pixels), with a sigmoid activa-
high component, which may make the representation dependent on the
tion. For the end-to-en, the Adam optimizer was implemented with a
observations in the dataset, rather than solely on the characterization
learning rate of 0.000001, during 2000 epochs, and a batch size of 16
of the image geometries themselves.
images. Fig. 7 shows saliency maps for three different compositions standing
Once the architecture was trained, embedded vectors of 64 scalar out the most salient regions, coded into compact descriptor, to carry
values were obtained, which are taken as compact descriptors related out the reconstruction. These reconstruction maps were taken from
to the structure of the observed images. These embedded vectors were mean vector embedding, which provides a closer approximation to
derived from the response vector of the means and variances, and the reconstruction. In these images we can see that within the set of
also the latent vector z. These embedded vectors were projected to dimensions of the embedded vector, there are elements that contain
a low-dimensional space under a supervised scheme using the LDA. features corresponding to different components of the nanofoam, such
Explainability maps were also obtained as a tool for observational as pores and matrix.
inspection of the main regions and structural features recovering 64 Besides, from set of variance embedded vectors, we analyzed
saliency maps generated in response to the mean vector. whether the spatial distribution of the vectors contains the visual

4
W.D. Romero et al. Materials Today Communications 37 (2023) 107407

Fig. 5. Representations of each type of embedded vector (𝜇, 𝜎, 𝑧), of each NMFs class (MF1, MF2 and MF3) and the geometric space in both training (right column) and evaluation
(left column) are shown for PCA.

Fig. 6. Representations of each type of embedded vector (𝜇, 𝜎, 𝑧), of each NMFs class (MF1, MF2 and MF3) and the geometric space in both training (right column) and evaluation
(left column) are shown for LDA.

Fig. 7. Grad-Cam activations for different dimensions of the 𝜇 embedding for each class of NMFs.

5
W.D. Romero et al. Materials Today Communications 37 (2023) 107407

Table 2
Machine learning models used for the classification of embedded vectors from different VAE architectures and three conventional architectures with the same
configuration. In this case, 𝜇 corresponds to the mean vector, 𝜎 to the variance vector, and 𝑧 is the linear combination of the two. These three vectors are
generated by the VAE coding architecture. In this case, conv refers to the embedded vector obtained from a conventional autoencoder with the same layer
configuration.
Approach 128 256 512
𝜇 𝜎 z conv 𝜇 𝜎 z conv 𝜇 𝜎 z conv
Decision Trees 0.82 0.9 0.59 0.81 0.8 0.87 0.52 0.79 0.8 0.86 0.44 0.8
KNN 0.66 0.92 0.66 0.44 0.65 0.8 0.63 0.52 0.62 0.69 0.58 0.49
Random Forest 0.88 0.98 0.86 0.83 0.87 0.94 0.76 0.93 0.87 0.92 0.63 0.83

Fig. 8. Distribution over the 2D latent space of the set of variance-embedded vectors for the test data set. Some of the data at the extremes of each class were also selected to
observe the distribution of the nanofoam characteristics.

characteristics of the NMFs within the latent space, as well as the concerning different copper–nickel compositions. Certainly, the num-
possibility of creating artificial images of new nanofoam configurations ber of types of NMF prepared was small to initially explore the effect of
from the reconstruction of new embedded vectors. Fig. 8 shows the nickel content on copper NMF, considering that both elements exhibit
distribution of the embedded vectors of the two-dimensional variance unlimited solubility when forming an alloy. From the findings achieved
(PCA) for the dataset of testing. From this dataset, we project images in this specific experimental setup, it is expected to carry out further
corresponding to some of the vectors into the space, thus observing a experimental analysis of other NMF types, with a larger number of
transformation in the size and number of pores as we move over the observational samples, to measure the characterization capability of the
proposed methodology. As an alternative tool, explainability maps were
latent space.
also obtained, which can provide evidence of the geometrical compo-
nents encoded in the embedded vector. Thus, the autoencoder focuses
6. Discussion and concluding remarks features such as pore spacing and pore geometry, characteristics used
by the experts during the analysis of nanofoams. Likewise, the analysis
This work introduced a variational autoencoder to describe ge- carried out from microstructures of porous electrodes allows for the
ometrical patterns corresponding to nanostructured porous metallic understanding of mechanical properties from different compositions.
foams with three different Cu/Ni ratios. The autoencoder was trained Such analyses are far from being definitive criteria for transferring this
following an image reconstruction task and succeeded in encoding technology in large-scale implementations. Nonetheless, these findings
embedded vectors with the ability to discriminate between geometries may be predominant in practical engineering applications. Recently,
of three different nanofoam compositions. some efforts have been introduced in the literature, describing for
Furthermore, through observations using a confocal microscope, we instance, how to fabricate Ni electrodes by DHBT [23].
codified nanostructured porous metallic foam patterns as embedded Future works should include the exploration of new computational
descriptors to represent the geometrical features. These descriptors architectures that can represent geometric characteristics and discrimi-
nate between different chemical compositions of the material. We also
may constitute an alternative characterization of metallic copper–nickel
intend to investigate the image analysis with observations in other
nanofoams, in this specific application, without involving tedious anal-
magnifications so that the computational tools can meet experimental
ysis and under-simplified pore measurements. Besides, these descrip-
requirements that can be costly or time-consuming in the synthesis
tors have discrimination capabilities from machine learning strategies, of the NMFs. It should be noted that reported results were obtained
achieving accuracy levels above 90%. The variance vector monitors the on a specific and limited dataset of copper–nickel nanofoams, explor-
dispersion of the reconstruction, which is indicative in the classification ing structural information observed from confocal images. However,
but is not consistent with the reconstruction information. the achieved results show potential applicability to support metallic
Additionally, the results showed that experimental environments nanofoams analysis, and a major exhaustive study should be carried out
can be simulated for the construction of NMFs. This fact holds great with a larger dataset to establish significant and statistical differences
potential for experts in the field to validate the geometrical features, among different compositions.

6
W.D. Romero et al. Materials Today Communications 37 (2023) 107407

CRediT authorship contribution statement [5] H.C. Shin, J. Dong, M. Liu, Nanoporous structures prepared by an electrochemical
deposition process, Adv. Mater. 15 (19) (2003) 1610–1614, http://dx.doi.org/10.
1002/adma.200305160.
William D. Romero: Methodology, Software, Visualization, Writing
[6] M. Ge, F. Su, Z. Zhao, D. Su, Deep learning analysis on microscopic imaging in
– review & editing. Santiago Torres-Bermudez: Data curation, Visu- materials science, Mater. Today Nano 11 (2020) 100087.
alization, Writing – review & editing. Brayan Valenzuela: Software, [7] K. Hiraide, K. Hirayama, K. Endo, M. Muramatsu, Application of deep learning
Supervision, Writing – review & editing. Cristian C. Viáfara: Investi- to inverse design of phase separation structure in polymer alloy, Comput. Mater.
gation, Conceptualization, Methodology, Supervision, Funding acquisi- Sci. 190 (2021) 110278.
[8] S.Y. Chung, C. Lehmann, M. Abd Elrahman, D. Stephan, Pore characteristics
tion, Writing – review & editing. Angel M. Meléndez: Investigation, and their effects on the material properties of foamed concrete evaluated using
Conceptualization, Methodology, Supervision, Funding acquisition, Re- micro-CT images and numerical approaches, Appl. Sci. 7 (6) (2017) 550.
sources, Writing – review & editing. Fabio Martínez: Investigation, [9] R. Liu, A. Antoniou, A relationship between the geometrical structure of a
Conceptualization, Methodology, Supervision, Funding acquisition, Re- nanoporous metal foam and its modulus, Acta Mater. 61 (7) (2013) 2390–2402.
[10] D. Oppedisano, Electrodeposited Metal Nanostructures Fabricated by Dynamic
sources, Writing – review & editing.
Hydrogen Bubble Templating (Ph.D. thesis), RMIT University, 2016.
[11] B.J. Plowman, L.A. Jones, S.K. Bhargava, Building with bubbles: the formation
Declaration of competing interest of high surface area honeycomb-like films via hydrogen bubble templated
electrodeposition, Chem. Commun. 51 (21) (2015) 4331–4346.
The authors declare that they have no known competing finan- [12] L. Jiao, J. Zhao, A survey on the new generation of deep learning in image
processing, IEEE Access 7 (2019) 172231–172263, http://dx.doi.org/10.1109/
cial interests or personal relationships that could have appeared to ACCESS.2019.2956508.
influence the work reported in this paper. [13] D.P. Kingma, M. Welling, et al., An introduction to variational autoencoders,
Found. Trends® Mach. Learn. 12 (4) (2019) 307–392.
Data availability [14] U. Michelucci, An introduction to autoencoders, 2022, arXiv preprint arXiv:
2201.03898.
[15] G.P. Way, C.S. Greene, Extracting a biologically relevant latent space from
Data will be made available on request. cancer transcriptomes with variational autoencoders, in: Pacific Symposium on
Biocomputing 2018: Proceedings of the Pacific Symposium, World Scientific,
Acknowledgments 2018, pp. 80–91.
[16] D. O’Malley, J.K. Golden, V.V. Vesselinov, Learning to regularize with a varia-
tional autoencoder for hydrologic inverse analysis, 2019, http://dx.doi.org/10.
This work was financially supported by Universidad Industrial de
48550/ARXIV.1906.02401.
Santander (Project 3733). [17] C. Doersch, Tutorial on variational autoencoders, 2016, arXiv preprint arXiv:
1606.05908.
References [18] D.P. Kingma, M. Welling, Auto-encoding variational bayes, 2013, arXiv preprint
arXiv:1312.6114.
[19] C. Vogelsanger, C. Federau, Latent space analysis of vae and intro-vae applied
[1] X. Ma, Z. Jing, C. Feng, M. Qiao, D. Xu, Research and development progress
to 3-dimensional mr brain volumes of multiple sclerosis, leukoencephalopathy,
of porous foam-based electrodes in advanced electrochemical energy storage
and healthy patients, 2021, arXiv preprint arXiv:2101.06772.
devices: A critical review, Renew. Sustain. Energy Rev. 173 (2023) 113111.
[20] T. Kurita, Principal component analysis (PCA), Comput. Vision: A Ref. Guide
[2] O.V. Kharissova, L.M. Torres-Martínez, B.I. Kharisov, Handbook of Nanomaterials
(2019) 1–4.
and Nanocomposites for Energy and Environmental Applications, Springer, 2021.
[21] A.M. Martinez, A.C. Kak, Pca versus lda, IEEE Trans. Pattern Anal. Mach. Intell.
[3] S. Vesztergom, A. Dutta, M. Rahaman, K. Kiran, I. Zelocualtecatl Montiel, P.
23 (2) (2001) 228–233.
Broekmann, Hydrogen bubble templated metal foams as efficient catalysts of
[22] W. Liu, R. Li, M. Zheng, S. Karanam, Z. Wu, B. Bhanu, R.J. Radke, O. Camps,
CO2 electroreduction, ChemCatChem 13 (4) (2021) 1039–1058.
Towards visually explaining variational autoencoders, in: Proceedings of the
[4] D. Nam, R. Kim, D. Han, J. Kim, H. Kwon, Effects of (NH4)2so4 and BTA on the
IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp.
nanostructure of copper foam prepared by electrodeposition, Electrochim. Acta
8642–8651.
56 (25) (2011) 9397–9405, http://dx.doi.org/10.1016/j.electacta.2011.08.025.
[23] D. Guay, J. Gaudet, M. Hao, V. Charbonneau, S. Garbarino, S.J. Thorpe, P.H.a.
Sobrinho, Porous ni electrodes and a method of fabrication thereof, 2021, Google
Patents, US Patent App. 16/870, 886.

You might also like