Hund’s matrix: A structure-based descriptor to predict the magnetic properties

Abstract

We propose a structure-based descriptor, where we have used Hund’s rules to represent the constituent atoms of crystalline materials. Coordinating atoms are used to create a local environment in the form of a matrix named Hund’s matrix. A descriptor is illustrated by training and testing the model using a random forest algorithm on various datasets to predict the formation energy and magnetization of Manganese based materials with good accuracy. Our results indicate that this descriptor is not only computationally fast, but it can capture the magnetic ordering of the solids by giving magnetization on the diverse datasets more accurately. We also demonstrated that the accuracy of the machine learning model depends on the correctness of the training dataset.