Machine Learned Potentials for Reaction Simulations

Machine learned potentials are widely used to describe the potential energy surface of near-equilibrium systems. To accurately explore and predict potential energy surfaces in the region of transition states requires active learning methods. In Li el al. [1], we show how machine-learned potentials can be used to simulate photochemical reactions, which are widely used by academic and industrial researchers to construct complex molecular compounds via mechanisms that often require harsh reaction conditions. Photodynamics simulations provide time-resolved insight into atomistic processes which are required to understand and predict reactivities and chemoselectivities. However, such simulations require thousands of costly quantum mechanical calculations per trajectory, which limits simulations to a picosecond time scale for most organic photochemical reactions. Westermayr et al. introduced a neural network-based method to accelerate the predictions of electronic properties and pushed the simulation limit to 1 ns for a methylammonium cation model system. We have adapted this methodology to develop two software tools, the Python Rapid Artificial Intelligence Ab Initio Molecular Dynamics (PyRAI2MD) tool for active learning-based molecular dynamics simulations, and the NNs4MD library which implements a single interface to a wide range of machine learning methods to be used in molecular dynamics simulations. Using these tools, we studied the cis-trans isomerization of trans-hexafluoro-2-butene and the 4pi-electrocyclic ring-closing of a norbornyl hexacyclodiene. We performed a 10 ns simulation for trans-hexafluoro-2-butene in just 2 days and discovered unknown reaction pathways in the ring-closing reaction. The same simulation would take approximately 58 years with traditional multiconfigurational photodynamics simulations. We are currently implementing graph neural networks in NNs4MD, in order to move from molecule-specific models to generalizable potentials.<br/><br/>[1] Li, J., Reiser, P., Boswell, B.R., Eberhard, A., Burns, N.Z., Friederich, P. and Lopez, S.A., 2021. Automatic discovery of photoisomerization mechanisms with nanosecond machine learning photodynamics simulations. Chemical science, 12(14), pp.5302-5314.