Active learning for predicting the enthalpy of mixing in binary liquids based on ab initio molecular dynamics

The enthalpy of mixing in the liquid phase is an important property for predicting phase formation in alloys. In multicomponent metallic liquids, it can be estimated from the binary interactions using a geometrical model, but data are available in less than a third of the binary systems. The prediction of this property in binary liquids is therefore important, and machine learning has recently achieved the highest accuracy. Further improvements requires acquiring high-quality data in liquids where models are poorly constrained. In this study, we propose an active learning approach to identify in which liquids additional data are most needed to improve an initial dataset that covers over 400 binary liquids. We identify a critical need for new data on liquids containing refractory elements, which we address by performing ab initio molecular dynamics simulations for 29 equimolar alloys of Ir, Os, Re and W. This enables more accurate predictions of the enthalpy of mixing, and we discuss the trends obtained for refractory elements of period 6. We use clustering analysis to interpret the results of active learning and to explore how our features can be linked to Miedema’s semi-empirical theory.