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

TL;DR

This paper develops and compares 3D CNN models for predicting elastic constants in nanoporous metals, demonstrating transfer learning and application to large-scale design optimization.

Contribution

It introduces a transfer learning approach with 3D CNNs for elastic modulus prediction in nanoporous metals, outperforming traditional descriptor-based models.

Findings

3D CNNs achieved R^2 up to 0.955 in elastic modulus prediction.

Transfer learning enabled effective model fine-tuning on smaller datasets.

Model applied to evaluate 100,000 nanoporous gold structures for optimal design.

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 ($R^2 = 0.704$), with to-performing DenseNet-201 architecture achieving $R^2 = 0.955$. 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.