Machine learning potentials and excited-state molecular dynamics

Artificial neural networks can learn the relationship between the nuclear geometry of a molecule
and the corresponding potential energy and, in this way, serve as a highly accurate and extremely
fast tool for predicting potential energy surfaces [1]. As an application, the simulation of an organic
reaction with such a machine learning algorithm is presented [2]. Furthermore, with the right neural
network setup, the chemical locality is inherently exploited, where the local region of a large
molecule is only weakly influenced by the atoms that are far from the region of interest [3]. As a
result, a vast amount of computational time can be saved when generating the so-called reference
data from which the neural network learns the shape of the potential energy surfaces. Such
improvements will enable us to perform on-the-fly dynamics calculations of large systems and long
time scales in the future.
Currently, we still use highly accurate but time-consuming ab initio calculations in semiclassical
molecular dynamics. Here, we study molecules in their excited electronic states by the surface
hopping method, where the nuclei move classically on potentials computed by quantum mechanics.
We have modified the surface hopping approach to incorporate not only kinetic dynamical
couplings but also any other arbitrary coupling in our so-called SHARC (surface hopping including
arbitrary couplings) method [4, 5]. The corresponding code is also publicly available [6]. Especially
spin-orbit couplings can now be treated in on-the-fly simulations. The method is applied to a variety
of systems, showing e.g. that intersystem crossing can take place on a femtosecond timescale even
in organic molecules with only relatively small spin-orbit couplings (see [5] and references therein).
Left: Artificial neural networks are algorithms inspired by the network of neurons in the human brain. Right: Logo of
the SHARC (Surface Hopping including ARbitrary Couplings) software.
References
[1] J. Behler, J. Phys.: Condens. Matter, 26 (2014) 183001.
[2] M. Gastegger, P. Marquetand J. Chem. Theory Comput., 11 (2015) 2187-2198.
[3] M. Gastegger, C. Kauffmann, J. Behler, P. Marquetand, J. Chem. Phys., 144 (2016) 194110.
[4] M. Richter, P. Marquetand, J. González-Vázquez, I. Sola, L. González, J. Chem. Theory
Comput., 7 (2011) 1253-1258.
[5] S. Mai, P. Marquetand, L. González, Int. J. Quantum Chem., 115 (2015) 1215-1231.
[6] S. Mai, M. Richter, M. Ruckenbauer, M. Oppel, P. Marquetand, L. González, SHARC
Program Package for Non-Adiabatic Dynamics, sharc-md.org (2014).