Physics-Informed and Data-Driven Molecular Dynamics Simulations of Iron under Extreme Conditions

We present a new Spectral Neighbor Analysis Potential (SNAP) machine-learning potential for large-scale molecular dynamics simulations of Iron. SNAP is a classical interatomic potential that expresses the energy of each atom as a linear function of selected bispectrum components of the neighbor atoms. The development of the SNAP potential entails three steps: (1) the creation of a training database comprised of a consistent and meaningful set of first-principles Density Functional Theory (DFT) data for Iron at a range of high pressures (0-400 GPa) and temperatures (0-6500 K); (2) the robust and physically guided training of the SNAP hyper-parameters based on DFT data using statistical data analysis; and (3) the validation of the SNAP potential in molecular dynamics simulations of Iron by evaluating transport properties at extreme conditions up to those prevalent in Earth's core.