Berlin 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 62: Developement of Calculation Methods III
MM 62.7: Talk
Thursday, March 21, 2024, 17:15–17:30, C 264
Warm Dense Hydrogen as a Benchmark for Machine-Learning Potentials — Bastian Jäckl1, Thomas Bischoff1, and •Matthias Rupp1,2 — 1University of Konstanz, Germany — 2Luxembourg Institute of Science and Technology, Luxembourg
Machine-learning potentials (MLPs) are fast data-driven surrogate models of potential energy surfaces that can accelerate ab-initio dynamics simulations by several orders of magnitude. The performance of MLPs is commonly measured as the prediction error in energies and forces on data not used for training. While low prediction errors on a test set are necessary, they are not sufficient for good performance in dynamics simulations. The latter requires physically motivated performance measures obtained from running accelerated simulations. The adoption of such measures, however, has been limited by the effort and domain knowledge required to calculate and interpret them. To overcome this limitation, we present data and scripts to automatically quantify the performance of MLPs in dynamics simulations of hydrogen under pressure. For this challenging benchmark system, we provide geometries, energies, forces, and stresses, calculated at the density functional level of theory for different temperatures and mass densities. We also provide scripts to automatically calculate, quantitatively compare, and visualize pressures, diffusion coefficients, stable molecular fractions, and radial distribution functions. Employing our benchmark, we show that several state-of-the-art MLPs fail to reproduce a crucial liquid-liquid phase transition, despite low test set errors in energies and forces.
Keywords: machine learning; inter-atomic potentials; hydrogen; molecular dynamics; benchmarks