Machine learning prediction of magnetic properties of Fe-based metallic glasses considering local structures

TL;DR

This study develops machine learning models to accurately predict the magnetic properties of Fe-based metallic glasses, considering complex local structures and multiple influencing factors, surpassing traditional physical models.

Contribution

Introduces ML models trained on experimental data that incorporate local structural effects, improving prediction accuracy for Fe-based metallic glasses' magnetism.

Findings

XGBoost and ANN models achieved R^2 >= 0.903

ML models effectively incorporate 13 influencing factors

Local structure significantly impacts magnetic property predictions

Abstract

Magnetism prediction is of great significance for Fe-based metallic glasses (FeMGs), which have shown great commercial value. Theories or models established based on condensed matter physics exhibit several exceptions and limited accuracy. In this work, machine learning (ML) models learned from a large amount of experimental data were trained based on eXtreme gradient boosting (XGBoost), artificial neural networks (ANN), and random forest to predict the magnetic properties of FeMGs. The XGBoost and ANN models exhibited comparably excellent predictive performance, with R^2 >= 0.903, mean absolute percentage error (MAPE) <= 6.17, and root mean squared error (RMSE) <= 0.098. The trained ML models aggregate the influence of 13 factors, which is difficult to achieve in traditional physical models. The influence of local structure, which was represented by the experimental parameter of the supercooled liquid region, presented a significant impact on the predictive performance of ML models. The developed ML-based method here can predict the magnetic properties of FeMGs by considering multiple factors simultaneously, including complex local structures.