Accelerating phase prediction via CALPHAD-informed machine learning and data augmentation

The immense compositional space of alloys presents a significant challenge for rapid phase prediction and renders exhaustive experimental exploration infeasible. This work introduces a machine learning (ML) framework that leverages CALPHAD-based thermodynamic modeling to accelerate phase stability predictions. The methodology is demonstrated through FCC phase prediction in a High Entropy Alloy system as a representative case study, with transferability further illustrated through BCC phase prediction using the same database and pipeline. A dataset of ∼72,000 compositions within a ten-element compositional space was generated using Thermo-Calc software. To mitigate data scarcity and class imbalance, a novel multidimensional interpolation strategy, grounded in the lever rule, was employed for data augmentation near phase boundaries, substantially expanded the training set without requiring additional CALPHAD calculations. Four ML models were evaluated: KNN, RF, GB and SVM, with the SVM consistently exhibiting superior performance across all configurations. Data augmentation improved balanced accuracy by 1–3%, with gains of up to 6.5 percentage points for the minority single-phase FCC class. The proposed methodology provides a robust, scalable, and computationally efficient strategy for the data-driven design of alloys.