3D Deep Learning with voxelized atomic configurations for modeling atomistic potentials in complex solid-solution alloys

Rahul Singh, Aayush Sharma, Onur Rauf Bingol, Aditya Balu, Ganesh Balasubramanian, Duane D. Johnson, Soumik SarkararXiv 2018 · arXiv

Introduction

High-Entropy Alloys (HEAs) represent a significant advancement in materials science, offering exceptional mechanical properties and thermal stability by combining five or more elements in near-equiatomic proportions. However, the vast compositional space of HEAs presents a formidable challenge for traditional computational modeling. Density Functional Theory (DFT) provides high accuracy but is computationally prohibitive for large systems, while empirical potentials like the Embedded Atom Method (EAM) often fail to capture the complex multi-element interactions inherent in these "cocktail" alloys. To bridge this gap, researchers have developed a 3D Deep Learning framework designed to predict potential energy surfaces with DFT-level accuracy at a fraction of the cost.

Methodology: The ML-PES Framework

The proposed framework, termed ML-PES (Machine-Learning-based Potential Energy Surface), treats atomic configurations as 3D spatial data. The core of this approach is the transformation of discrete atomic coordinates into a continuous 3D grid, a process known as voxelization.

The ML-PES workflow: atomic configurations are voxelized and processed through a 3D CNN to predict directional forces.

Voxelization and Bounding Boxes

In the voxelization step, each voxel is assigned a structural descriptor ($V$) based on atomic numbers and local densities, modeled using Gaussian distributions. This representation allows the model to capture the spatial context of the chemical environment. The researchers explored two bounding box strategies for defining the grid:

  • Internal Bounding Box (ibb): A dynamic grid that scales to fit the atoms in each specific snapshot.
  • Outer Bounding Box (obb): A fixed grid size determined by the maximum displacements across the entire dataset.

3D CNN Architecture

The voxelized data is processed by a 3D Convolutional Neural Network (CNN). Unlike standard fully connected (dense) networks, 3D CNNs are inherently suited for spatial data, as they utilize local filters to identify geometric patterns and neighbor interactions—the physical drivers of atomic forces.

Data and Training

The model was trained on an equiatomic MoTaW (Molybdenum, Tantalum, and Tungsten) alloy, a representative complex solid-solution system. The dataset comprised 70,000 snapshots generated via Ab-Initio Molecular Dynamics (AIMD) at temperatures ranging from 100K to 1500K. This wide temperature range ensured the inclusion of diverse atomic displacements, providing a robust training set for the neural network.

Results and Discussion

A primary finding of the study was the superior performance of the fixed Outer Bounding Box (obb) strategy. While the flexible "ibb" approach might seem more adaptive, the "obb" method preserves absolute scale information, which is critical for accurate force prediction.

Validation loss comparison showing that the obb strategy significantly outperforms both the ibb and traditional dense neural networks.

The "obb" methodology achieved a validation loss of approximately 0.0075 at a $16 \times 16 \times 16$ resolution, a significant improvement over the ~0.02 loss observed in traditional dense networks. Furthermore, the model demonstrated high fidelity in predicting directional forces (e.g., $F_x$), with predicted medians and quantiles closely matching the ground-truth DFT data.

The study also identified a "resolution sweet spot." While higher voxel resolutions generally improve accuracy, the gains diminish beyond $16 \times 16 \times 16$, suggesting that this resolution provides an optimal balance between predictive accuracy and computational overhead.

Conclusion and Future Directions

The 3D CNN framework proves to be a powerful surrogate for expensive first-principles calculations in complex alloy systems. By effectively learning local chemical environments through voxelization, ML-PES enables the simulation of larger systems over longer timescales.

Future work will focus on scaling the model to handle alloys with more than three elements and addressing slight variances in force predictions across different spatial axes ($F_y$ and $F_z$). Additionally, the researchers plan to integrate ML-PES into full-scale Molecular Dynamics simulations to validate bulk material properties, such as thermal conductivity and elasticity, against experimental benchmarks.

Summary updated

Abstract

The need for advanced materials has led to the development of complex, multi-component alloys or solid-solution alloys. These materials have shown exceptional properties like strength, toughness, ductility, electrical and electronic properties. Current development of such material systems are hindered by expensive experiments and computationally demanding first-principles simulations. Atomistic simulations can provide reasonable insights on properties in such material systems. However, the issue of designing robust potentials still exists. In this paper, we explore a deep convolutional neural-network based approach to develop the atomistic potential for such complex alloys to investigate materials for insights into controlling properties. In the present work, we propose a voxel representation of the atomic configuration of a cell and design a 3D convolutional neural network to learn the interaction of the atoms. Our results highlight the performance of the 3D convolutional neural network and its efficacy in machine-learning the atomistic potential. We also explore the role of voxel resolution and provide insights into the two bounding box methodologies implemented for voxelization.
BibTeX
@misc{singh2018d,
  author = {Rahul Singh and Aayush Sharma and Onur Rauf Bingol and Aditya Balu and Ganesh Balasubramanian and Duane D. Johnson and Soumik Sarkar},
  title = {{3D Deep Learning with voxelized atomic configurations for modeling atomistic potentials in complex solid-solution alloys}},
  year = {2018},
  eprint = {1811.09724},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1811.09724},
}
Figure