Mapping high entropy state spaces for novel material discovery

Abstract

High-entropy alloys show promising properties for novel catalytic designs, but their vast potential configurations make them challenging to study computationally. Additionally, the traditional methods for data acquisition required to train neural networks on these broad systems can be inefficient. To address this, we propose an active learning methodology that integrates genetic algorithms with deep convolutional neural networks trained on Density Functional Theory calculations via a simple closed feedback loop. This approach streamlines data acquisition and the exploration of large configurational spaces simultaneously. We illustrate its effectiveness on high-entropy clusters of variable sizes and compositions, the vast state spaces of which are automatically explored and trained on, so as to generate and predict the stability of any cluster within the latent space given minimal computational requirements. Importantly, this method is adaptable for use in a variety of other systems of different sizes, chemical compositions, and stoichiometry.