Abstract

The accurate modelling of chemical reactions in solution is a well-established problem in chemistry. Ab initio methods provide accurate descriptions of bond-making and bond-breaking events, but the high computational costs prohibit an explicit treatment of solvent particles over chemically meaningful length and timescales. Recent developments in machine learning have alleviated some of these issues, yet several challenges remain, such as the accurate modelling of reactions at interfaces or under spatially anisotropic conditions.

In this talk, I will lay out ways to address these problems through the development of a robust training procedure for generating neural network potentials. Our protocol makes use of several important features, such as active learning, priority training, dataset reduction, and corrections in the long-range interaction. Using the example of the hydration reaction of CO2 , we demonstrate the effectiveness of our protocol in exploring chemistry under gaseous, bulk, and interfacial conditions using just a single model. Our results show that the solvation environment can significantly impact the energy landscape of the system, affecting reaction barriers, product energies and stabilities, mechanisms of reactions, and conformational stability.

We believe that our protocol can be applied to more complicated schemes of reaction, such as pericyclic reactions, ‘on-water’ reactions, and liquid-phase heterogeneous catalysis. Our work contributes to the development of more efficient and accurate methods for modelling chemical reactions under realistic conditions, with potential applications in fields such as materials science, drug design, and environmental chemistry.