Transferable 3D Convolutional Neural Networks for Elastic Constants Prediction in Nanoporous Metals

Sergei Zorkaltsev, Rafał Topolnicki, Tal-El Carmon, Santhosh Mathesan, Paweł Dłotko, Dan Mordehai, Maciej Haranczyk · 2026

Read Transferable 3D Convolutional Neural Networks for Elastic Constants Prediction in Nanoporous Metals by Sergei Zorkaltsev, Rafał Topolnicki, Tal-El Carmon, Santhosh Mathesan, Paweł Dłotko, Dan Mordehai, Maciej Haranczyk (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-20890.txt.

Introduction

Sergei Zorkaltsev and six colleagues compare shared 3D convolutional networks for elastic constants prediction in porous metals. Metal shape strongly shapes how they respond to load. The authors generated 6,000 gold and 422 silver porous structures, computing three elastic modulus parts with molecular dynamics, for 19,263 data points. They compare a fully connected network trained on pre-made topological markers against several 3D convolutional designs from machine vision. The 3D networks beat the marker baseline, whose R-squared was 0.704, with top DenseNet-201 reaching 0.955.

Catalog record

CREATORSergei Zorkaltsev, Rafał Topolnicki, Tal-El Carmon, Santhosh Mathesan, Paweł Dłotko, Dan Mordehai, Maciej Haranczyk
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE74942
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.20890

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Transferable 3D Convolutional Neural Networks for Elastic Constants Prediction in Nanoporous Metals Sergei Zorkaltsev, Rafał Topolnicki, Tal-El Carmon, Santhosh Mathesan, Paweł Dłotko, Dan Mordehai, Maciej Haranczyk Abstract The topology of nanoporous metals is crucial for determining their mechanical response. In this work, we generated 6,000 gold and 422 silver nanoporous structures and calculated three components of elastic modulus with Molecular Dynamics simulations, resulting in 19,263 data points. This study compared two distinct approaches of predicting elastic modulus: a Fully-Connected neural network trained on precomputed topological descriptors, and several 3D Convolutional neural network architectures adapted from computer vision. The 3D CNNs outperformed the descriptor‐based baseline model ( R 2 = 0.704 R^{2}=0.704 ), with top-performing DenseNet‐201 architecture achieving R 2 = 0.955 R^{2}=0.955 . Additionally, the effects of training grid resolution, dataset size, and descriptor integration into a model were investigated. We further demonstrated model robustness through Transfer learning: a pretrained model was fine-tuned on a much smaller dataset of denser gold structures and the dataset of denser silver structures. Finally, the trained model was employed to evaluate the mechanical properties of 100,000 stochastic nanoporous gold structures and identify the Pareto optimal designs. keywords Nanoporous structures, Convolutional neural network, Elastic Constants † † equal-contributors: These authors contributed equally to this work. † † equal-contributors: These authors contributed equally to this work. † † equal-contributors: These authors contributed equally to this work. 1 Introduction Nanoporous metals (such as noble: Au, Ag, Pd, Pt and non-noble: Al, Ni, Cu, Zn) have found a wide range of applications in many fields: chemical catalysis 17 ; 60 ; 39 , sensors 28 ; 56 , and energy storage 4 ; 13 because of their high conductivity, high surface-to-volume ratio, and catalytic activity. Depending on desired mechanical properties, porosity, and application, these nanoporous structures can be manufactured by de-alloying 47 ; 43 , chemical deposition 2 , template fabrication 29 , and other methods. Existing approaches to determining the relationship between the microstructure and mechanical properties of nanoporous metals are mostly based on scaling laws. One of these scaling laws is the Gibson-Ashby model 8 , which correlates the elastic modulus and the yield stress of the structure with its solid fraction, or the Roberts-Garboczi model 41 , which corrects the results for small solid fractions considering a percolation threshold. Although accurate in predicting properties of some types of materials, these scaling laws can deviate substantially for nanoporous metals 59 , lattices 67 , and triply periodic minimal surface (TPMS) structures 6 . For instance, in the study of Peng et al. 36 , researchers manufactured nanoporous silver samples by joining silver nanowires, reaching porous structure with high solid fractions (above 50%), for which the Gibson-Ashby model is believed to be accurate. However, they found out that the yield strength of nanosilver porous structure is higher than predicted by the Gibson-Ashby model, reaching 2.6 GPa. To explain this, they have shown that the yield strength also depends on the thickness and length of the nanowires, obeying a smaller-is-stronger law. On the other hand, experimental studies have shown that the Gibson–Ashby scaling law remains valid for nanoporous structures when size effects and network connectivity or topological parameters are appropriately considered in predicting their mechanical properties 26 ; 63 . In addition to experimental approaches, simulations enable a more detailed analysis of topology and its relationship to mechanical properties. These studies facilitate the examination and modification of the existing scaling laws by incorporating new topologically related parameters or proposing novel equations. Some finite element method (FEM) studies pointed out that the effective Young’s modulus of nanoporous structures also depends on the scaled genus, proposing either a linear 30 ; 31 or a quadratic 51 relationship. The precision of topology-independent scaling laws deteriorates in nanoporous structures in confined volume, as was demonstrated in nanoporous gold (NPG) pillars 62 , as the number of ligaments that support the deformation is limited. The need to account for the load-bearing ligaments was previously demonstrated in molecular dynamics (MD) simulations 32 . In these compression simulations, it was shown that the number of load-bearing ligaments can differentiate the Young’s moduli of NPG nanopillars of the same solid fraction but with different topologies. Accordingly, MD simulations can serve as a powerful tool to find structure-property correlations in nanoporous structures, since elastic properties are well captured by interatomic potentials and the simulation cells usually include only a few load-bearing ligaments, which makes the results more sensitive to topology. Despite this, the relations between the structures and elastic properties of nanoporous stochastic structures remain unresolved. As an alternative to experiments and simulations, deep learning (DL) has demonstrated great success in many research areas, and it can be used to predict properties directly on input structures. The popularity of DL models comes from a specific type of neural network called Convolutional Neural Networks (CNNs). Initially developed for 2D data and particularly Computer Vision related problems, the architectures like AlexNet 21 , VGG 50 , ResNet 11 , and their modifications are still applied for many tasks, including materials science. For instance, Kato et al. 19 trained a CNN to classify microstructure images of samples produced from two different alloys with two different heat treatment conditions. Their results show that with only 140 images available for training, it is possible to achieve near-ideal classification accuracy with data augmentation. In the work of Wu et al. 61 authors developed a similar CNN-based approach for the regression task of permeability prediction. The neural network architecture of this study was a physics-informed CNN, which takes both 2D images and physical parameters (porosity and surface area ratio) as the inputs. This approach allowed for achieving high accuracy of predictions while being orders of magnitude faster than conventional methods for permeability estimation. It is also possible to apply convolutions to three-dimensional data by 3D-CNNs. Cawte and Bazylak 3 trained a 3D-CNN to predict the permeability of numerically generated gas diffusion layer materials. A relatively simple and straightforward CNN architecture of this work showed a low prediction error with a mean 5-fold R^{2} score of 0.9885. The authors concluded that the performance of this simple deep learning model is comparable to the best available models for permeability prediction. Rao and Liu 38 applied 3D-CNN to a dataset of generated composite materials with random spherical inclusions. The mechanical properties of these materials were estimated with FEM simulations and used as targets for model training. Their results showed a high accuracy of 3D-CNN predictions, lower computational costs compared to conventional methods, and the potential to transfer learning, which allows training on a smaller dataset. Santos et al. 44 developed a multiscale 3D-CNN model to predict fluid flow through a porous media. This multi-scale approach combined several subsequent networks, each of them being aware of previous predictions and having a finer size of convolutional filter. The model was trained on a mix of synthetic and real data and then tested on vuggy core samples, realistic sandstone samples, and micro-CT images of fractured carbonates. Other examples of tasks solved with the application of 3D-CNN include natural convection in porous materials 55 , up-scaling of rock micro-CT images 42 , detection of clogging in microfluidic systems 65 , and

Read the complete text