PhaseForge: Integrating Machine-Learned Potentials with CALPHAD for Rapid Alloy Phase Diagram Construction

The construction of CALPHAD thermodynamic databases remains a major bottleneck in Integrated Computational Materials Engineering (ICME) workflows, traditionally requiring expensive density functional theory (DFT) calculations and extensive experimental measurements. We present PhaseForge, an open-source framework that accelerates this process by replacing DFT with machine learning interatomic potentials (MLIPs). PhaseForge automates an end-to-end pipeline: structure sampling via special quasirandom structures (SQS) through ATAT, MLIP-based energy calculations, phonon and molecular dynamics simulations for temperature-dependent thermodynamic contributions, and fitting of CALPHAD-compatible thermodynamic database (TDB) files. The framework supports multiple universal MLIPs—including GRACE, ORB, CHGNet, and SevenNet, among others—enabling systematic benchmarking against DFT and experiment. We demonstrate PhaseForge through the construction of binary phase diagrams, benchmarking the resulting phase boundaries against established CALPHAD assessments. Results show that modern MLIPs reproduce DFT-level thermodynamic properties at a fraction of the computational cost, enabling rapid exploration of alloy composition spaces otherwise prohibitive with first-principles methods. By dramatically reducing the time and cost of thermodynamic database development, PhaseForge represents a step toward autonomous, high-throughput ICME workflows.