(203g) Training Lean: New Approaches for Developing Machine Learning Interatomic Potentials for Simulations of Chemical Reactivity in Liquids

Machine-learning interatomic potentials (MLIPs) promise large-scale molecular dynamics simulations with first-principles accuracy. They can be used to understand molecular reactions in complex systems, such as green, non-aqueous solvents with tunable, functional properties. However, the generation of MLIPs to study reactions in solution requires extensive datasets as non-local solvent reorganization must be included in the training. While these datasets can be obtainable through umbrella sampling or other enhanced sampling methods, they are on the order of tens of thousands of frames, making the training of MLIP for chemical reactions expensive.

Here, we describe our recent work which introduces a coarse-sampling approach to deploy MLIPs in chemical reactions in solution, requiring only hundreds of training frames. [1] We study the room-temperature decomposition of a 2:1 molar ratio of ethylene glycol: choline chloride via S_N2 reaction, in a generalizable workflow which can be extended to other organic solvents. Our workflow also enables mechanistic understanding, as we find from our coarse-sampling approach, that fluctuations in the hydrogen bonds bind chlorine in high-energy states, promoting the uphill reaction. We offer design rules for green solvents, recommending that future design of these solvents should explicitly consider the hydrogen bond network as a strong perturbation of the potential energy surface which alters solvent properties such as reactivity. In summary, our general approach can be used to efficiently study a variety of neotoric solvents undergoing chemical reactions.

[1] Julia H. Yang, Amanda Whai Shin Ooi, Zachary A.H. Goodwin, Yu Xie, Jingxuan Ding, Stefano Falletta, Ah-Hyung Alissa Park, Boris Kozinsky. Room-temperature decomposition of the ethaline deep eutectic solvent, The Journal of Physical Chemistry Letters, in press.
DOI: 10.1021/acs.jpclett.4c03645