Leveraging Fragment-Based Representations in Active Learning and Reinforcement Learning Frameworks for Materials Design

This talk will focus on the development and application of two types of methods for accelerating materials design. First, we will focus on developing reinforcement learning methods that are used to accelerate the design of functional materials, including radical-based polymers and organic optoelectronic materials. In our first demonstration of the reinforcement learning scheme, we show that this framework can integrate with quantum chemistry calculations in real-time, and through a careful design of the learning curriculum, we are able to find a diverse set of molecules with desired singlet and triplet energy levels. Second, we will describe our work on developing representations for predicting the polymer physics of disordered polymers at a much lower computational cost than current coarse-grained methods. One advantage of the new representation is that it avoids specifying the longest length of the chain in advance. In addition, this representation works well with a set of highly charged sequences, uncovering new insights to the fundamental interactions and scaling behavior of these systems. We will then discuss the compatibility of this representation with reinforcement learning methods.