A simple feature enriched CGCNN for predicting magnetization in transition metal compounds

Abstract

Accurate prediction of magnetic properties is essential for accelerating the discovery and modeling of magnetic materials. While machine learning methods such as Orbital Field Matrix capture ferromagnetic (FM) systems fairly well, there is a lack of approaches tailored for ferrimagnetic (FiM) compounds, whose complex yet technologically vital behavior remains largely unexplored. To address this, we extend the Crystal Graph Convolutional Neural Network (CGCNN) by integrating atomic spin magnetic moments as node attributes and structural parameters as edge attributes. This enhancement enables the network to outperform existing methods for FM materials and effectively model magnetism in FiM compounds. We utilize the Materials Project database to curate datasets comprising 3d transition-metal (TM) compounds for the training and evaluation of the model. The trained model generalizes effectively to unseen complex systems and demonstrates strong transferability across experimental and computational datasets of TM and rare-earth compounds. In materials families like Heusler alloys and MXenes, accuracy improves with minimal representative data during training. This enables accurate predictions for novel and unique magnetic compounds, even with limited datasets.