Active-learning-based efficient prediction of ab-initio atomic energy: a case study on a Fe random grain boundary model with millions of atoms

TL;DR

This paper introduces an active learning method that efficiently predicts atomic energies in large-scale grain boundary models with DFT accuracy, significantly reducing computational costs for massive atomic systems.

Contribution

The study presents a novel active learning approach using uncertainty reduction and replica DFT atomic energy to accurately predict atomic energies in models with millions of atoms.

Findings

Prediction error decreases faster than random sampling.

Method is suitable for massively parallel computing environments.

Effective for modeling large grain boundary systems with high accuracy.

Abstract

We have developed a method that can analyze large random grain boundary (GB) models with the accuracy of density functional theory (DFT) calculations using active learning. It is assumed that the atomic energy is represented by the linear regression of the atomic structural descriptor. The atomic energy is obtained through DFT calculations using a small cell extracted from a huge GB model, called replica DFT atomic energy. The uncertainty reduction (UR) approach in active learning is used to efficiently collect the training data for the atomic energy. In this approach, atomic energy is not required to search for candidate points; therefore, sequential DFT calculations are not required. This approach is suitable for massively parallel computers that can execute a large number of jobs simultaneously. In this study, we demonstrate the prediction of the atomic energy of a Fe random GB model containing one million atoms using the UR approach and show that the prediction error decreases more rapidly compared with random sampling. We conclude that the UR approach with replica DFT atomic energy is useful for modeling huge GBs and will be essential for modeling other structural defects.