Refining Machine Learning Potentials through Thermodynamic Theory of Phase Transitions
The DiffTTC method uses top-down learning to fine-tune machine learning potentials by directly correcting predicted phase transition temperatures against experimental data. By employing Differentiable Trajectory Reweighting, the approach minimizes free energy differences between phases at target conditions. This successfully corrected the phase diagram of pure Titanium up to 5 GPa, matching experimental reference within tenths of a Kelvin.