Machine learning assisted phase diagram prediction of refractory multi-component alloys

The design of refractory high-entropy alloys (RHEAs) requires accurate knowledge of phase stability across temperatures, yet existing machine learning (ML) models are constrained by small, imbalanced datasets limited to room-temperature predictions. This work introduces an ML framework for the prediction of complete phase diagrams across composition and temperature space. By extracting 129,339 data points from calculated ternary phase diagrams for eight refractory elements (Cr, Mo, Nb, Ta, Ti, V, W, Zr) from 300 K to 3000 K, we transform the complex problem of phase diagram prediction into a series of binary classification tasks for individual phases. A neural network (NN) model trained on this data achieves 95% test accuracy. Crucially, the model's capability extends beyond point-based phase prediction to reconstructing full isothermal sections of ternary and higher-order systems. The model's accuracy in predicting the phase diagram is rigorously validated against established experimental and calculated phase diagrams. This approach is powerfully demonstrated by systematically screening all ternary systems and by constructing pseudo-binary phase diagrams for complex quaternary and quinary alloys to identify those exhibiting a single BCC phase across a wide temperature range. This study presents a reliable, data-driven framework for phase diagram prediction in multicomponent systems and for accelerating the design of novel refractory alloys with tailored phase stability.