Condensed Matter Physics Seminar: Florian Knoop

Abstract:

I present a scheme that couples density functional theory (DFT) simulations, self-consistent phonon theory in the temperature-dependent effective potentials (TDEP) framework [1], and machine-learning potentials to perform ab initio materials simulations at finite temperature for a variety of observables at the cost of <1000 supercell calculations. These range from thermodynamic properties of superconducting high-pressure hydrides [2] to spectroscopic observables under realistic conditions, in particular Raman response in simple and complex systems such as perovskites for solar cell applications [3]. The scheme is accelerated by a recent equivariant neural network potential [4, 5]. This furthermore enables non-perturbative simulations of thermal transport via the Green-Kubo approach [5, 6], delivering the accuracy of its fully ab initio counterpart [7] at a fraction of the computational costs.

[1] https://tdep-developers.github.io/tdep/
[2] D Laniel, F Trybel, B Winkler, FK, et al., Nat Commun 13, 6987 (2022)
[3] S. Filippone, et al., Phys Rev Mater 4, 091601 (2020); N. Benshalom, et al., Phys Rev Mater 6, 033607 (2022)
[4] JT Frank, OT Unke, KR Müller, in Advances in Neural Information Processing Systems, Vol. 35 (2022), pp. 29400–29413
[5] MF Langer, JT Frank, FK, Arxiv 2305.01401
[6] MF Langer, FK, et al., Arxiv 2303.14434
[7] FK, M Scheffler, C Carbogno, Phys Rev B 107, 224304 (2023); FK, et al., Phys Rev Lett 130, 236301 (2023)