Machine learning of phase diagrams

Abstract

By starting from experimental- and ab initio-determined phase diagrams (PDs) of materials, a machine learning (ML) method is developed to infer the free energy function for each phase. The ML method is based on a custom two-step explore-exploit k-nearest neighbor strategy, which samples the multidimensional space of Gibbs free energy parameters and user-defined physical constraints into a database of millions of PDs in order to identify the target material properties. The method presented herein is 1000× to 100 000× faster than currently available approaches, and defines a new paradigm on the quantification of properties of materials and devices. As an example application, the developed methodology is combined with the most widely used thermodynamic models – the regular solution, Redlich–Kister, and sublattice formalisms – to infer the properties of materials for lithium-ion battery applications in a matter of hours, reconstructing without human bias, well-established CALPHAD formulations while identifying previously missed stable and metastable phases and associated properties. For the EC–DMC–PC systems, the ML method allows to distinguish between stable and metastable phase boundaries, while simultaneously considering the relevant phases. For the high-power density LiFePO₄ chemistry, a room-temperature metastable phase is identified. Its appearance highlights a previously unreported driving force for the transformation kinetics between lithiated and delithiated states that serves as a stepping stone to access a high-temperature eutectoid state that can be applied to engineer solid-state chemistries. For the high-energy density LiCoO₂ chemistry, a highly lithiated electronically insulating phase is thermochemically favorable, particularly at grain corners and boundaries, greatly improving the description of the experimental voltage profile, irrespective of the used baseline free energy model to describe the relevant phases.