Neural networks meet non-covalent interactions: Chalcogen bonding in a solution
Reliable description of competing non-covalent interaction represents a long-term goal of quantum chemistry. To overcome the shortcomings of static computations in implicit solvent, it is necessary to run molecular dynamics (MD) simulations of explicitly solvated system. However, ab initio MD is often prohibitively expensive for large scale simulations.
Herein we present a simulation protocol based on Neural Network Potentials (NNPs)  that reproduce the hybrid DFT energies and forces. While this approach has been increasingly used for studying solid state systems with low chemical complexity, its application for solvated molecular systems with many elements remains scarce.  In this work, we study the complex of 4,5,6,7-tetrafluorobenzo-2,1,3-telluradiazole and Cl- solvated in tetrahydrofuran. This system contains two competing weak interactions, a chalcogen bond and anion-π, which are important in the supramolecular chemistry, organocatalysis and anion recognition.
We train two sets of NNPs: one using the periodic GFN0-xTB semiempirical Hamiltonian  as a cheap baseline potential, and second directly reproducing the DFT data. We observe a significant improvement of the free energy profile compared to the Amoeba polarizable force field. Moreover, we show that the baselined NNPs are more robust than the direct NNPs. Our approach demonstrates the importance of accurate DFT-based potentials and enhanced sampling techniques in the description of weakly bonded systems in solution.
 Jörg Behler, Michele Parrinello, Physical Review Letters, 2007, 98, 146401.
 Kevin Rossi, Veronika Jurásková., Raphael Wischert, Laurent Garel, Clémence Corminboeuf, Michele Ceriotti, Journal of Chemical Theory and Computation, 2020, 16, 8, 5139–5149.
 Christoph Bannwarth, Eike Caldeweyher, Sebastian Ehlert, Andreas Hansen, Philipp Pracht, Jakob Seibert, Sebastian Spicher, Stefan Grimme, WIREs Computational Molecular Science, 2021, 11, 1493.