Temperature- and ab initio functional-dependent machine learning interatomic potentials for high-pressure materials

The recent accelerated development of machine learning interatomic potential (MLIP) algorithms has enabled quantum-accurate simulations at experimental time and length scales. However, applying these methods to calculate high-pressure phase diagrams remains challenging due to the high computational cost of obtaining temperature-dependent MLIPs or the need to train MLIPs beyond standard LDA/GGA levels. Here, we present low-cost and efficient procedures to develop such MLIPs. We demonstrate this approach on several metallic and semiconducting materials, where traditional MLIP methods fail to produce accurate simulations.