Can Machine Learning Predict Solvation Effects on Energies and Geometries of Highly Charged Molecules?

Recent machine-learning models have shown that implicit solvation based on continuum solvation models can be learned efficiently with graph neural networks. However, previous studies have focused almost exclusively on neutral or only weakly charged molecules, leaving the strongly charged regime essentially unexplored. Here, we systematically investigate which model architectures and architectural components are required to predict solvation energies and corresponding forces for molecules with total charges ranging from -5 to +5. Using reference data generated from the conductor-like screening model for realistic solvation (COSMO-RS) for water, acetonitrile, and cyclohexane, we show that an explicit and robust treatment of molecular charge is essential for reliable performance across charge states. At the same time, highly charged systems prove easier to learn than neutral and singly charged ones, whose solvation energies depend more strongly on subtle structural effects. Finally, we demonstrate that the machine-learning models enable stable solvent-aware geometry optimization and can be used for efficient conformational sampling in solution.