Thermodynamic Integration By Neural Network Potentials Based On First-Principles Dynamic Calculations

Simulation-size effect in evaluating the melting temperature of material is studied systematically by combining thermodynamic integration (TI) based on first-principles molecular-dynamics (FPMD) simulations and machine learning. Since the numerical integration to determine the free energies of two different phases as a function of temperature is very time consuming, the FPMD-based TI method has only been applied to small systems, i.e., less than 100 atoms. To accelerate the numerical integration, we here construct an interatomic potential based on the artificial neural-network (ANN) method, which retains the first-principles accuracy at a significantly lower computational cost. The free energies of the solid and liquid phases of rubidium are accurately obtained by the ANN potential, where its weight parameters are optimized to reproduce FPMD results. The ANN results reveal a significant size dependence up to 500 atoms.