Application of batch learning for boosting high-throughput ab initio success rates and reducing computational effort required using data-driven processes

TL;DR

This paper introduces a machine learning-enhanced method to optimize input parameters in high-throughput ab initio calculations, significantly increasing success rates and reducing computational costs in material studies.

Contribution

It presents a systematic, data-driven approach integrating machine learning into high-throughput workflows to improve ab initio calculation success and efficiency.

Findings

Success rate increased from 64.8% to 94.3%.

Computational effort per successful relaxation reduced by 17%.

Applicable to large-scale magnetic multilayer studies.

Abstract

The increased availability of computing time, in recent years, allows for systematic high-throughput studies of material classes with the purpose of both screening for materials with remarkable properties and understanding how structural configuration and material composition affect macroscopic attributes manifestation. However, when conducting systematic high-throughput studies, the individual ab initio calculations' success depends on the quality of the chosen input quantities. On a large scale, improving input parameters by trial and error is neither efficient nor systematic. We present a systematic, high-throughput compatible, and machine learning-based approach to improve the input parameters optimized during a DFT computation or workflow. This approach of integrating machine learning into a typical high-throughput workflow demonstrates the advantages and necessary considerations for a systematic study of magnetic multilayers of 3$d$ transition metal layers on FCC noble metal substrates. For 6660 film systems, we were able to improve the overall success rate of our high-throughput FLAPW-based structural relaxations from $64.8 \%$ to $94.3\ \%$ while at the same time requiring $17\ \%$ less computational time for each successful relaxation.