Exploring Molecules with Low Viscosity: Using Physics-Based Simulations and De Novo Design by Applying Reinforcement Learning

Abstract

Molecules with viscosities lower than those of conventional organic solvents are highly sought after for applications in electrochemical devices such as batteries and capacitors. These molecules improve the electrical resistance of devices, enhancing their efficiency, especially at low temperatures. To identify new molecules with low viscosities, we conducted extensive molecular dynamics (MD) simulations on 10,000 molecules selected from the GDB-17 chemical structure database, specifically choosing molecules with fewer than 12 heavy atoms. Additionally, we performed density functional theory (DFT) calculations to determine the energies of the highest occupied molecular orbitals (HOMO) of these molecules as a surrogate for the oxidation potential. We used the data on viscosity and HOMO levels as training sets to develop machine-learning models that predict these properties. Using these models, we carried out molecular de novo design using the REINVENT method, a reinforcement-learning approach utilizing SMILES strings. This method aimed to identify molecules that minimize viscosity while maintaining sufficiently low HOMO levels for stability. The approach successfully identified new chemical structures with viscosities below 2 mPa·s and suitably low HOMO energies. We synthesized a novel compound from the top candidates and validated our predictions experimentally. The experimental results closely matched our predictions, demonstrating that combining physics-based simulations with reinforcement learning is an effective strategy for designing novel molecules with targeted properties.