Temperature dependent interatomic machine-learning potentials of metallic systems

In the last decades, the development of machine-learning interatomic potentials (ML-IAP) has been one of the most successful directions in application of artificial intelligent methods in physics, chemistry and material science. To date, quantum-accuracy of ML-IAPs is often claimed by training a single transferable temperature-independent ML-IAP to the ground state energies, forces and/or stresses of the prepared configurations, achieving small error metrics of these quantities and reproducing a few material properties of interest. In this talk, we show that this approach is often insufficient when constructing a temperature-independent ML-IAP for metallic systems. We introduce an efficient two-step approach to generate temperature dependence. We demonstrate that the two-step approach works well on metallic liquid systems with different temperature-dependance of their electronic and structural properties.