Complete machine learning description of chemical reactions in solution

The projection of a 3N dimensional space onto a low dimensional collective variable (CV) space is one of the bottlenecks of the study of physical transformations. Many machine learning (ML) schemes have been proposed to devise an optimal CV using classical forcefields. These kinds of ML methods are however out of reach for ab initio simulations that would require millions of CPU.h only to produce the training data. Even with an optimal CV, complete ab initio studies of physical transformations are very demanding in computational time, this problem can be solved using machine learning potentials (MLP). Here, we propose to combine a MLP method devised in the team along with a machine learning CV to accurately study the properties of a benchmark chemical reaction in solution with ab initio accuracy and state of the art CV.