Can simple exchange heuristics guide us in predicting magnetic properties of solids?

The Kanamori-Goodenough-Anderson rules are a textbook heuristic for predicting magnetism. They connect bond angles to magnetic ordering for some transition metal compounds. Such domain knowledge is of high importance for building predictive machine learning models in scenarios with scarce data. Yet, there has been no statistical, large-scale evaluation of the heuristic. Here, we evaluate this heuristic on an experimental database of magnetic structures. We observe that the heuristic is largely satisfied, and we discuss the exceptions. We then demonstrate how integrating this heuristic into machine learning models for predicting magnetic ordering enhances prediction quality. Notably, these magnetism models are also capable of predicting if non-collinear magnetic ordering might occur. Furthermore, the heuristic provides a useful benchmark for evaluating theoretical methods that calculate magnetic properties.