Berlin 2024 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 11: Data Driven Material Science: Big Data and Workflows II

MM 11.2: Talk

Monday, March 18, 2024, 16:00–16:15, C 243

From ab-initio to scattering experiments using neuroevolution potentials — •Eric Lindgren¹, Adam Jackson², Zheyong Fan³, Christian Müller⁴, Jan Swenson¹, Thomas Holm-Rod⁵, and Paul Erhart¹ — ¹Department of Physics, Chalmers University of Technology, Gothenburg, Sweden — ²Centre for Sustainable Chemical Technologies and Department of Chemistry, University of Bath, United Kingdom — ³College of Physical Science and Technology, Bohai University, Jinzhou, People's Republic of China — ⁴Department of Chemistry and Chemical Engineering, Chalmers University of Technology, Gothenburg, Sweden — ⁵ESS Data Management and Software Center, Copenhagen, Denmark

Machine-learned interaction potentials have in recent years emerged as an appealing alternative to traditional methods for obtaining forces for molecular dynamics simulations, combining the computational efficiency of semi-empiricial potentials with the accuracy of ab-inito methods. In particular, Neuroevolution potential (NEP) models, as implemented in the GPUMD package, are highly accurate and computationally efficient, enabling large scale MD simulations with system sizes up to millions of atoms with ab-initio level accuracy. In this work, we present a workflow for constructing and sampling NEPs using the `calorine` package, and how the resulting trajectories can be analysed with the `dynasor` package to predict observables from scattering experiments. We focus on our recent work on crystalline benzene as an example system, but the approach is readily extendable to other systems.

Keywords: machine learning; force fields; potentials; molecular dynamics; neural network