Acceleration of phase diagram construction by machine learning incorporating Gibbs' phase rule

Abstract To efficiently construct phase diagrams of alloy systems, a machine learning-based method advanced by thermodynamics on phase equilibria is proposed. With the use of uncertainty sampling in active learning, the next point to be synthesized or measured can be recommended to efficiently draw the phase diagram. For appropriate recommendations, two ingenuities are introduced in the machine learning method: training data preparation when the multiphase coexisting region is detected and search space reduction based on the Gibbs’ phase rule. We demonstrate the construction of ternary phase diagrams using our machine learning method by incorporating these ingenuities. The complicated phase diagram of alloy systems could be effectively plotted even when knowing only the information of single-component systems in the initial step. The recommendation made by our machine learning method can help reduce the number of experiments required to construct a phase diagram to approximately 1/8 compared with random sampling.