Virtual screening now plays an instrumental role in accelerating the discovery of new functional materials. Some notable recent applications include screenings that have found new candidates for organic photovoltaics, organic light-emitting diodes, and organic redox flow battery electrolytes. In these investigations, vast combinatorial libraries of molecules (on the order of 10<sup>6</sup>) were generated and optimal properties were screened for using quantum chemistry methods on a significant fraction of the library—a computationally expensive process. Recently, several different machine learning models have been developed that generate new molecules. In this work, we explore the implementation of generative models that use reinforcement learning from quantum chemistry calculations to guide the optimization of organic molecules with desired properties. Properties of interest include bandgaps and reduction potentials, which are difficult to calculate accurately with only cheminformatics or machine-learning-based approaches. However, they can be captured using quantum chemistry methods. We apply the method to generate new high and low potential molecules for organic redox flow batteries. This application of machine learning driven virtual screening exploits the strengths of both generative models and virtual screening. Details on both implementation and challenges in optimizing the feedback between these two modules are extensively explored.

Coupling generative models to virtual screening: Application to organic redox flow batteries

255th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.