Machine Learning Many-Body Interactions for Charged Colloidal Suspensions

Abstract

Suspensions of charged colloidal particles are complex systems of mesoscopic colloidal particles in an electrolyte solution composed of microscopic water molecules and ions. An accurate description of these systems is vital for predicting their properties and the stability of the suspension. Direct computer simulations can provide us with accurate predictions, but are too computationally expensive for large systems due to the huge amount of microscopic water molecules and ions. Direct simulation of the entire system is also inefficient, as our interest lies primarily in the behavior of the mesoscopic colloids. We employ machine learning methods to learn effective many-body interactions for a colloid-only system from direct simulations of the full system. The accuracy of this machine learning approach is compared to established effective interactions such as Poisson-Boltzmann and DLVO theory, and its efficiency against the direct simulation of the full colloidal suspension.