Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 36: Data Driven Materials Science: Big Data and Work Flows – Microstructure-Property-Relationships (joint session MM/CPP)
MM 36.7: Vortrag
Donnerstag, 30. März 2023, 12:00–12:15, SCH A 251
Stress and Heat Flux via Automatic Differentiation — •Marcel F. Langer1,2,3, Florian Knoop3,4, J. Thorben Frank1,2, Christian Carbogno3, Matthias Scheffler3, and Matthias Rupp3,5 — 1BIFOLD – Berlin Institute for the Foundations of Learning and Data, Berlin, Germany — 2Machine Learning Group, Technische Universität Berlin, Germany — 3The NOMAD Laboratory at the Fritz Haber Institute of the Max Planck Society and Humboldt University, Berlin, Germany — 4Theoretical Physics Division, Department of Physics, Chemistry and Biology (IFM), Linköping University, Sweden — 5Materials Research and Technology Department, Luxembourg Institute of Science and Technology (LIST), Luxembourg
Computationally efficient approximations of the Born-Oppenheimer potential energy surface can be obtained by parametrising an analytical force field based on a set of reference calculations. Inspired by recent developments in machine learning, such potentials can include equivariant semi-local interactions through message-passing mechanisms and rely on automatic differentiation (AD), overcoming the need for manual derivative implementations or finite-difference schemes. We provide a unified framework for using AD in such state-of-the-art potentials, and discuss how AD can be used to efficiently and simply compute the stress tensor and the heat flux. We validate the framework by predicting thermal conductivity for selected semiconductors and insulators with an equivariant machine learning potential [1].
[1]: J.T. Frank, O.T. Unke, K.-R. Müller, arXiv 2205.14276 (2022).