Interpretable machine learning for atomic scale magnetic anisotropy in quantum materials

Abstract

The rising demand for digital storage and environmental concerns necessitate ultra-high-density, energy-efficient solutions. Atomic-scale magnets (ASMs) based on transition metal (TM) dimers on defective graphene exhibit promising magnetic anisotropy energy (MAE) values, providing a robust barrier against magnetization reversal. However, identifying optimal TM-substrate configurations is challenging when relying solely on density functional theory (DFT) calculations with spin-orbit coupling. To address this, we developed a machine learning (ML) model trained on scalar-relativistic DFT data using a tree-based gradient boosting approach. Our model implicitly captures key physical interactions from second-order perturbation theory, ensuring reliable MAE predictions for systems beyond the training set. By bridging computational efficiency with interpretability, this work contributes to the development of ASMs for spintronics and quantum materials, offering a pathway to next-generation data storage technologies.