Prediction of Exchange Bias for Magnetic Heterostructure Nanoparticles with Machine Learning

Abstract

Exchange bias is essential for the stability and control of nanoparticles’ magnetic properties for their application as rare-earth-free permanent magnet, magnetic storage, magnetic hyperthermia, and catalysis. Core–shell structures of magnetic bimagnetic particles have garnered increasing interest due to their larger coercive and exchange bias fields, tunable blocking temperatures, and enhanced Neel temperature. However, the design approach of nanoparticles with exchange bias using a computational method has a high computational cost and offers limited efficiency in predicting complex core–shell nanoparticle systems. Machine learning (ML) predictions provide a transformative approach to the design and optimization of materials with desirable exchange bias (EB) properties by offering rapid and precise evaluations of material compositions and configurations. This study addresses this gap by developing an ML model for the prediction of EB field in magnetic nanoparticles, which provide a fast and effective alternative for traditional computational methods. Hence, comparative analysis of ML models, including the Kolmogorov-Arnold Network (KAN), is for predicting EB in core–shell and heterostructure nanoparticles. Among the predictive models, XGBoost demonstrated superior performance, achieving R² values of 0.74 and 0.75 on the test and validation data sets, respectively. KAN showed reduced generalization power with the R² test of 0.67 but was more accurate in predicting high values of the EB. The Shapley additive explanation (SHAP) analysis revealed unexpected dependencies between nanoparticle properties and magnetic behavior, offering new insights for optimizing a material design. These findings are highly relevant for developing materials for rare-earth-free permanent magnets, magnetic storage, magnetic hyperthermia, and catalysis, where precise control of magnetic properties is crucial.