Abstract

Photodynamic simulations are powerful tools to decipher the fundamental mechanisms of light-initiated reactions. However, high computational costs underlying calculations of accurate ab initio potentials for excited states limit the applicability of this method to short time scales, and the possible breaking and formation of bonds after light excitation complicates the search for a suitable reference method [1].

In this talk, I will present our efforts to improve photodynamics simulations with machine learning (ML) models that can describe energies and forces of different spin multiplicities and couplings between them. Our method is termed the SchNarc approach [2] and provides efficient generation of training sets, as well as a phase-free training algorithm that allows fitting of excited state properties that are arbitrary with respect to their sign. The power of our method is demonstrated with long time scale photodynamics of the methyleneimmonium cation [3] and on the complex photochemistry of the amino acid tyrosine. The latter illustrates the expressive power of ML models that learn from different levels of theory and allow the dynamics of systems that could not be treated with the required accuracy otherwise [4].

[1] J. Westermayr, P. Marquetand Chem. Rev. 121(16), 9873-9926 (2021).

[2] J. Westermayr, M. Gastegger, P. Marquetand J. Phys. Chem. Lett. 11(10), 3828-3834 (2020).

[3] J. Westermayr, M. Gastegger, M. Menger, S. Mai, L. González, P. Marquetand Chem. Sci. 10, 8100-8107 (2019).

[4] J. Westermayr, M. Gastegger, D. Vörös, L. Panzenböck, F. Jörg, L. González, P. Marquetand, arXiv: 2108.04373 (2021).