Abstract

Neural networks can efficiently calculate atomistic potential energy surfaces, allowing for large-scale molecular dynamics simulations.

In this talk I’ll describe how so-called high-dimensional neural network potentials, as proposed by Behler and Parrinello [1], can be constructed and fitted to reproduce ab-initio or density functional theory results. I’ll also demonstrate some recent applications of this approach to nuclear quantum effects in electrolyte solutions [2] and anisotropic proton diffusion at solid-liquid interfaces [3].

Moreover, I’ll briefly describe another type of neural network potential based on a graph convolutions, with a particular emphasis on “PiNet”, which we recently demonstrated could predict a wide range of properties for molecules, liquids, and materials [4].

J. Behler, M. Parrinello. Phys. Rev. Lett. 98, 146401
M. Hellstrom, M. Ceriotti, J. Behler. J. Phys. Chem. B 2018, 122, 44, 10158–10171
M. Hellstrom, V. Quaranta, J. Behler. Chem. Sci., 2019, 10, 1232-1243
Y. Shao, M. Hellstrom, P.D. Mitev, L. Knijff, C. Zhang. J. Chem. Inf. Model. 2020, 60, 3, 1184–1193