Multiobjective Design of Electrolyte Solvents via Physics-Based Modeling and Reinforcement Learning

Molecular modeling and machine learning are becoming essential tools in designing materials. Traditionally, materials are optimized for a single target property; however, this often results in a compromise on other critical properties. Therefore, it is crucial to develop methods that consider multiple target properties for practical material design in application-specific contexts. For instance, electrochemical devices such as lithium-ion batteries and various capacitors require molecules with lower viscosities, suitable electronic properties, and boiling points above the maximum operating temperature. This study demonstrates the use of physics-based modeling and advanced machine learning techniques to design molecules that meet such multiproperty criteria. Viscosity and boiling point are computed using molecular dynamics (MD), while HOMO energy (a surrogate for oxidative stability) is evaluated using density functional theory (DFT). We developed and validated two MD-based approaches for boiling point estimation by comparing with experimental data. Machine learning models are trained for each property and incorporated into a reinforcement learning-based framework for de novo molecular generation using SMILES strings. We explore two design criteria: (i) minimizing viscosity while ensuring a high boiling point, and (ii) minimizing both viscosity and HOMO energy under the same boiling point constraint. From the generated candidates, 200 top-performing molecules are selected and further validated using physics-based simulations. Our results demonstrate that a combination of physics-based models and reinforcement learning enables the discovery of molecules with tailored thermal, fluidic, and electronic properties, providing a practical strategy for multiproperty optimization in materials design.