Functional Dependence of Melting Behavior and Thermodynamic Properties of Silicon and their Application to Machine Learned Potentials

Accurate thermodynamic modeling of silicon’s phase diagram requires density functional theory (DFT)-level accuracy but predicted phase equilibria are highly sensitive to the choice of exchange–correlation functional. Here we use the ChIMES machine-learned interatomic model as a proxy to assess how DFT functionals, ranging from GGA to Hybrid-GGA, affect the predicted solid–solid and solid–liquid phase diagrams of silicon. Leveraging ChIMES’ efficiency, we map functional-dependent variations in melting and transition behavior that would be prohibitively expensive to evaluate directly with DFT methods. The results show significant shifts in predicted phase boundaries across functionals, emphasizing that functional selection strongly impacts the thermodynamic description of silicon and the transferability of machine- learned interatomic models for extreme-condition phase transformations. These findings provide guidance for constructing DFT training sets for machine-learned potentials and for interpreting simulation results derived from those training sets.

This work was performed in part under the auspices of the U.S. Department of Energy by Lawrence Livermore National. Laboratory under Contract DE-AC52-07NA27344.