A Machine Learning Approach for Increased Throughput of Density Functional Theory Substitutional Alloy Studies

TL;DR

This paper presents a machine learning method that significantly accelerates the exploration of substitutional alloy design spaces by predicting ion positions and energies, reducing computational costs in density functional theory studies.

Contribution

The study introduces a neural network approach to predict initial ion positions and energies, enabling faster alloy configuration evaluations without full relaxations.

Findings

Achieved up to 37x speedup in alloy configuration calculations

Reduced relaxation time by an average of 64 CPU-hours per calculation

Validated neural network predictions with low error margins compared to DFT results

Abstract

In this study, a machine learning-based technique is developed to reduce the computational cost required to explore large design spaces of substitutional alloys. The first advancement is based on a neural network approach to predict the initial position of ions for both minority and majority ions before ion relaxation. The second advancement is to allow the neural network to predict the total energy for every possibility minority ion position and select the most stable configuration in the absence of relaxing each trial position. In this study, a bismuth oxide materials system, (Bi$_{x}$La$_{y}$Yb$_{z}$)$_2$ MoO$_6$, is used as the model system to demonstrate the developed method and potential computational speedup. Comparing a brute force method that requires the calculation of every possible minority concentration location and subsequent relaxation there was a 1.3x speedup if the neural network (NN) was allowed to predict the initial position prior to relaxation. This speedup is a result in an average decrease of 4 wall hour (64 cpu-hrs) reduction in relaxation for individual calculations. Implementation of the second advancement allowed the NN to predict the total energy for all possible trials prior to relaxation, resulting in a speedup of approximately 37x. Validation was done by comparing both position and energy between the NN to DFT calculations. A maximum vector mean squared error (MSE) of 1.6x10$^{-2}$ and a maximum energy MSE of 2.3x10$^{-7}$ was predicted. This method demonstrates a significant computational speedup, which has the potential for significant computational savings for larger compositional design spaces.