Functional materials design often requires that the desired molecules (or materials) simultaneously satisfy multiple desired properties, such as electrochemical properties, stability, and synthetic accessibility. This talk will discuss several strategies we have developed for efficiently navigating chemical space and accelerating the inverse design of new functional organic molecules and materials, and the physical insights we gain during their design and deployment.
First, we will discuss our developments in utilizing reinforcement learning methods to design 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 reinforcement learning curriculum, we are able to find a rich set of molecular scaffolds with desired singlet and triplet electronic energy levels. We will discuss the importance of uncertainty quantification and the effects of external design constraints on the molecular design process, integrate more complex combinations of desired properties, and apply a similar framework to the problem of assigning the electronic spectra of conjugated polymers.
Second, we will describe our work on developing physically motivated representations for predicting the “top-line” polymer physics properties (e.g., radius of gyration) of sequence-controlled polymers at a much lower computational cost than coarse-grained methods. For this talk, we will discuss the logic behind the development of these representations, how to make the models stronger through careful design of the training and testing sets, and their utility in more complex tasks, such as the design of polymers with specified structural domains.
If time permits, we will discuss our work to understand the fundamental design principles for optimizing chemical reactions under external forces (mechanochemistry). Here, we use a combination of high-throughput screening, optimization methods, and graph-based neural network potentials to conduct a broad search for reactions that can be significantly accelerated by external forces that are achievable in modern mechanochemical reactors. Our methods use machine learning potentials in combination with reaction path searching protocols (e.g., nudged elastic band and the growing string methods) to find potential transition states. We then explore candidate “activatable” coordinates—specific deformation modes that lead to enhanced reaction rates—by analyzing a mix of localized and normal coordinates. The most promising reactions and degrees of motion are then verified by higher-level calculations.