Predicting hot-electron free energies from ground-state data

Machine-learning potentials are usually trained on the ground-state, Born-Oppenheimer energy surface, which depends exclusively atomic positions and not simulation temperature. This disregards effect of thermally-excited electrons, that is important in metals, essential to description warm dense matter. An accurate physical these effects requires nuclei move a temperature-dependent electronic free energy. We propose method obtain machine-learning predictions this at an arbitrary electron temperature using training data from ground-state calculations, avoiding need train potentials, benchmark it metallic liquid hydrogen conditions core gas giants brown dwarfs. work demonstrates advantages hybrid schemes use consideration combine predictions, providing blueprint for development similar approaches extend reach atomistic modelling by removing barrier between physics data-driven methodologies.