(483d) Deep Reinforcement Learning to Address Uncertainties in Computational Molecular Design

The tuning of microscale molecular properties to improve the functionality and efficiency of products and processes have long served as the primary means for the development of novel products and processes. However, in current paradigms, uncertainty estimates of both property models and property data are not factored into the design of molecules, leading to suboptimal or even infeasible solutions due to the potential violation of property, product, and/or process constraints. Herein, we present DRL-CMD, a deep reinforcement learning (DRL) framework for addressing the uncertainties present in computational molecular design (CMD) by providing molecular solutions with uncertainty estimates. A standard molecular design case study is used to compare the performance of our approach to published solutions in the literature. Computational results show that the DRL framework provides superior solutions that factor in uncertainties associated with molecular properties. The flexibility and efficiency of the devised DRL framework are promising for solving not only molecular design problems, but also multiscale product and process design problems.