Search for eutectic high entropy alloys by integrating high-throughput CALPHAD, machine learning and experiments

Abstract

We present a comprehensive study on the identification of eutectic high entropy alloys (EHEAs) through integration of CALculation of PHAse Diagrams (CALPHAD), machine learning (ML), and experimental data. By performing high-throughput CALPHAD calculations to obtain the temperature differences between liquidus and solidus phases (ΔT) and employing gradient descent optimization to identify local minima in ΔT surface, we obtained a reliable dataset for EHEAs across 5–6 component alloy families, which effectively addresses current limitations in both the quality and availability of EHEA data. In conjunction with literature-based experimental data, this dataset serves as the foundation for ML models trained with an XGBoost classifier. The physical descriptors with the most significant effects on the classification of eutectics and non-eutectics are identified. Our study reveals that configurational entropy alone yields a remarkable 98% classification accuracy, elucidating its dual role in phase stabilization and melting point depression. For the first time, an explicit phase selection rule to identify eutectics has been derived from an artificial neural network model, which facilitates efficiently screening EHEAs without resorting to CALPHAD nor ML models. This study presents a robust, data-driven strategy applicable not only to EHEAs but also to a broader range of alloy systems.