Thermodynamically Accurate Foundational Machine Learning Potentials

This proposal aims to enhance the predictive accuracy of the MACE-MP machine learning model for modeling metallic phase transitions, a crucial step in developing advanced materials. Therefore, the project will benchmark theMACE-MP-03b model on unary phase diagrams of 16 metals that are important in high-performance alloys for mechanical or catalytic applications. The team will refine the benchmark model for experimental phase diagrams by correcting predictions for selected solid-liquid and solid-solid phase transitions using our Diff TTC method. Finally, the team will validate the refined model to ensure reliable predictions of phase stability and transition temperatures. The resulting foundational model will be thermodynamically accurate, opening up the use of foundational models for the design and study of alloy materials across various technological fields.