Presentation Information
[9a-N304-5]A Machine Learning Compatible Workflow to Optimize Free Energy Models using Experimental Phase Equilibria Data
〇(P)Wenhao Zhang1, Yusuke Matsuoka1, Taichi Abe1 (1.NIMS)
Keywords:
Phase diagram,Optimization
Recent advances in universal machine learning interatomic potentials (UMLIPs) enable high-throughput prediction of phase stability and phase diagrams. However, UMLIPs results contain errors on the level of 1-2 kJ/mol, due to both machine learning and DFT training data. Thus, they cannot reliably predict phase boundaries and phase transitions, which are curcial to guide experimental effort. Inspired by CALPHAD thermodynamic assessment, where parameterized free-energy models are optimized against experimental phase equilibria, we develop a PyTorch-based, ML compatible framework for optimizing general thermodynamic models, including machine learning models, from experimental phase equilibrium data. By defining an efficient, differentiable loss function with respect to observed equilibria, gradient-based optimization can be applied. We expect such framework can be integrated with UMLIPs to provide better prediction of phase diagrams.