Machine Learning the Periodic Table

Machine learning models - from those predicting the interatomic potential to those estimating electronic and functional properties - have become an integral part of the atomistic modeling toolbox. The most established architectures treat individual elements as separate entities, learning their interactions independently, and leading to a steep scaling of the data requirements with the degree of chemical complexity. Several recent models, instead, incorporate the chemical nature of the atoms into fixed-length tokens, that are combined with structural information in a way that avoids a dramatic increase in complexity with growing numberd of species.<br/>I will discuss models that apply explicitly a reduction of dimensionality of chemical space, showing how this makes traditional ML potential capable of achieving accurate, transferable predictions for high-entropy alloys containing up to 25 d-block metals. The highly interpretable architecture provides insights into the functioning of more complicated deep-learning models, and makes it possible to tackle important applicative questions including the onset of short-range order in HEAs, and the preferential segregation of elements at surfaces, which is of great relevance for applications to catalysis.