Berlin 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
HL: Fachverband Halbleiterphysik
HL 28: Focus Session: Heat transport at the nanoscale: theory meets experiment
HL 28.8: Talk
Wednesday, March 20, 2024, 12:30–12:45, EW 203
Computing Green-Kubo Thermal Conductivities with Semi-Local Machine-Learning Potentials — •Marcel F. Langer1,2,3,4, Florian Knoop4,5, J. Thorben Frank2,3, Christian Carbogno4, Matthias Scheffler4, and Matthias Rupp6 — 1COSMO Laboratory, EPFL, Lausanne, Switzerland — 2BIFOLD, Berlin, Germany — 3ML Group, TU Berlin, Germany — 4NOMAD Laboratory at the FHI of the Max Planck Society and IRIS Adlershof of HU Berlin, Germany — 5Theoretical Physics Division, IFM, Linköping University, Sweden — 6Materials Research and Technology Dept., Luxembourg Institute of Science and Technology, Luxembourg
The Green-Kubo method is a rigorous framework for heat transport simulations in materials, but requires an accurate description of the potential-energy surface and converged statistics. In this context, machine-learning potentials can achieve the accuracy of first-principles methods while allowing to reach well beyond their simulation time and length scales at a fraction of the cost. Recently developed potentials can include equivariant semi-local interactions through message-passing mechanisms and use automatic differentiation to obtain derivatives. We explain how to define and efficiently implement the heat flux for such potentials [1]. Based on this, we present a framework for running GPU-accelerated Green-Kubo calculations with machine-learning potentials, and demonstrate its use through the calculation of the thermal conductivity of several solid semiconductors and insulators.
[1]: M.F. Langer et al., Phys. Rev. B 108, L100302 (2023); M.F. Langer et al., J. Chem. Phys. 159, 174105 (2023)
Keywords: machine learning; thermal transport; message-passing neural networks; automatic differentiation