Inverse organic materials design via integration of reinforcement learning with real-time quantum chemistry | Poster Board #439

Generative molecular design strategies have emerged as promising alternatives to library-based high-throughput screening approaches for multi-objective problems in large chemical spaces. To date, generative models with reinforcement learning approaches usually employ low-cost methods to evaluate the quality of the generated molecules (e.g., in drug discovery). However, for functional molecular materials tasks, such low-cost methods are either not available or would require the generation of large amounts of training data to train surrogate machine learning models.

Here, we present our continued developments to integrate reinforcement learning frameworks with real-time quantum chemistry calculations as part of the evaluation step. Our approach involves designing optimal curriculum strategies for the reinforcement learning framework and developing adaptive computational protocols that accelerate loops through the reinforcement learning cycle. We will first present our work on discovering a diverse set of molecular lead candidates for singlet fission and triplet-triplet applications, showing that the framework can be employed to discover a diverse set of promising lead molecules. These molecules can also be interpreted based on fundamental chemically motivated design rules.

We then present our latest developments on the methodology, including integrating the approach with alternative molecular representations to SMILES strings, developing adaptive strategies for switching between exploratory and exploitative searches, and integrating uncertainty quantification. For these new developments, we test the applications on the task of designing highly stable organic radical-based materials, which have potential applications in next-generation energy storage systems.