Machine Learning Model for Nd2Fe14B-Based Permanent Magnets

We demonstrate an efficient machine learning (ML) model for the prediction of magnetic property changes in Nd2Fe14B-based permanent magnets given a large range of different impurity elements. We show that relatively simple models can be sufficient to capture complex changes in the saturation magnetization Ms and the magnetocrystalline anisotropy constant K1. As the necessity for recycling the raw material of permanent magnets increases, the variety of impure chemicals and their concentrations increase as well. Some chemical elements with antiferromagnetic or complex magnetic ground states like Cr, Mn and Sm pose difficulties in the training of an ML model that can be effectively mitigated by feature engineering. This enables us to create a single model capable of describing more than twenty substitutional elements in a wide range of concentrations.