Coupling Reinforcement Learning Approaches with Efficient Quantum Chemical Models for Organic Electronic Materials Discovery and Design

In this talk, we will focus on two angles for accelerating organic materials design, with our application spaces focusing on optoelectronic and redox-active organic materials. The first angle focuses on developing discriminative machine learning models that can predict non-extensive organic material properties (e.g., redox potential, reorganization energy) from a low number of training points. In this area, our main areas of focus are on developing appropriate features for the model, quantifying uncertainty in prediction, and developing active learning strategies for developing robust models with as few training points as possible. The second angle focuses on developing methods that pick the right materials to simulate and ensuring that algorithms avoid the numerous local minima that exist in chemical space. We will discuss our work on leveraging reinforcement learning schemes for the inverse design of conjugated and non-conjugated materials and if time permits, our work on integrating these reinforcement learning methods with other representations for molecular materials.