(657j) Many-Body Machine Learning Force Fields for Explicit Solvent Modeling

Modeling solvation effects in computational chemistry often requires fully explicit treatments (i.e. molecular dynamics) to accurately and reliably capture solute-solvent interactions. Unfortunately, such methods require pre-parameterized force fields or many quantum mechanics calculations that can incur a prohibitively high computational cost. Recent developments in efficiently trained symmetric gradient-domain machine learning (sGDML) force fields have shown that they can enable molecular dynamics with CCSD(T) accuracy. However, a central limitation of sGDML force fields is their lack of transferability in that they are only applicable to systems the force field has been trained on. To address this issue, we consider the applicability of using a many-body decomposition comprised of multiple sGDML models that run simultaneously. Specifically, we trained sGDML models to learn energies for monomer, dimer, trimer, and tetramer interactions based on ab initio molecular dynamics simulations of solvent clusters. We found that many-body GDML models for water, acetonitrile, and methanol have more accurate force and energy prediction when compared to standard sGDML models. Furthermore, we demonstrate that force fields with up to 3-body interactions are reasonably accurate and transferable for dynamics simulations on larger solvent clusters of arbitrary size.