Abstract
<jats:p>Predicting alloy phase diagrams across broad chemical spaces remains difficult because conventional CALPHAD assessments are system-specific, while direct machine-learning models of phase boundaries do not necessarily correspond to a thermodynamically consistent free-energy description. Here, we develop a differentiable thermodynamic-learning approach for large-scale learning of alloy phase equilibria, demonstrated using binary liquidus prediction as a representative solid–liquid equilibrium problem. A neural network maps elemental and solid-hull descriptors to the coefficients of a compact Redlich–Kister liquid Gibbs-energy model. Liquidus temperatures are obtained through a common-tangent construction against a fixed coarse-grained solid-phase convex hull, allowing the model to be trained end to end directly from phase-equilibrium data. Using liquidus data for 286 binary systems and five-fold system-level cross-validation, the model reaches a final average test mean absolute error (MAE) of 129.9 K and a best-checkpoint average of 126.8 K. Under the same system-level splits, direct learning of independently fitted Redlich–Kister parameters gives a test MAE of 166.5 K, while a simple endpoint-interpolation baseline gives 233.5</jats:p>