Modeling Many-Body\nInteractions in Water with Gaussian\nProcess Regression

We report a first-principles water dimer potential that\ncaptures\nmany-body interactions through Gaussian process regression (GPR).\nModeling is upgraded from previous work by using a custom kernel function\nimplemented through the KeOps library, allowing for much larger GPR\nmodels to be constructed and interfaced with the next-generation machine\nlearning force field FFLUX. A new synthetic water dimer data set,\ncalled WD24, is used for model training. The resulting models can\npredict 90% of dimer geometries within chemical accuracy for a test\nset and in a simulation. The curvature of the potential energy surface\nis captured by the models, and a successful geometry optimization\nis completed with a total energy error of just 2.6 kJ mol<sup>–1</sup>, from a starting structure where water molecules are separated by\nnearly 4.3 Å. Dimeric modeling of a flexible, noncrystalline\nsystem with FFLUX is shown for the first time.