Fragmentation methods have been extensively employed in the literature over the past few decades to speed up quantum chemistry calculations. In recent years elaborate fitting procedures of many-body potential terms to reference “first principles” quantum calculations significantly decreased the computational cost of such fragmentation methods. This allowed the construction of very accurate potentials for molecular simulations like the highly successful MB-pol water model, 1,2 which combines a physically motivated representation of long-range interactions with accurate low-order terms of the many-body expansion. MB-pol accurately describes experimental properties of water from clusters to liquid phase. 1 Application of the underlying model to systems with molecular fragments other than water, however, is non-trivial and requires the development of new parameterizations. 3 A crucial point when developing new potentials for different systems is the generation of high-quality training sets. In here we employ the active learning procedure via query by committee as described by Smith et. al 4 for this task, using ANI-like neural networks 5 for one-body, two-body, and three-body potentials. The generated training sets are then fed into a permutationally invariant polynomial fitting infrastructure to generate “MB-pol”-like potentials for new systems. 3 By construction, the newly generated potentials accurately represent both short- and long-range interactions, which is important for a proper description across gas and liquid phases and interfaces. To demonstrate the power of this approach, we present results regarding the aqueous solvation of dinitrogen pentoxide (N2O5) and of hypochlorous acid (HOCl). The reactivity of N2O5 or HOCl with aerosols has been the focus of many studies for more than a decade. An important ingredient in understanding the reactive uptake of these pollutant gases by aerosol particles is their interaction with small water clusters and also solvation in bulk water. By using the newly generated potentials in biased MD simulations we are capable to obtain the free energy of solvation in order to predict the solubility of N2O5 or HOCl, a quantity that is currently not accessible experimentally for N2O5 but a key ingredient in atmospheric models of reactivity.

Developing accurate many-body potentials using machine learning for simulations of N2O5 and HOCl in water

Macromolecular Chemistry: The Second Century

[Author Benefits](https://scimeetings.acs.org/?utm_source=pubs_content_marketing&utm_medium=website&utm_campaign=0320_MCF_NPI_Launch_Spring_Homepage&ref=pubs_content_marketing)

[How to use SciMeetings](https://storage.googleapis.com/dl.morressier.com/acs/SciMeetings-ACSSpring2020-HowToGuide_M3020-Final.pdf)

[Frequently Asked Questions](https://storage.googleapis.com/dl.morressier.com/acs/SciMtng_ACSNMSp2020_FAQ_M3120_Final.pdf)

ACS Spring 2020 National Meeting & Expo

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.