This project will develop methods to fine-tune foundational machine-learning interatomic potentials using experimental phase diagram data. The approach aims to overcome known limitations of density functional theory for rare-earth and actinide materials, enabling more predictive atomic simulations of critical materials important to energy technologies and national security.
2026
Fine-tuning Atomic Models from Phase Diagrams to Go Beyond DFT
Preliminary D-DOS application to Actinides. Typical schemes a) collate DFT training data, select MLIP architecture then b) fit MLIP before sampling to evaluate free energies. c) D-DOS learns a descriptor entropy function from a single sampling campaign, returning a differentiable free energy predictor valid over a broad parameter range. d) shows preliminary actinide application, fine-tuning a MACE foundation model.
Other Researchers
Thomas Swinburne, Assistant Professor, Mechanical Engineering