|
ChemGymRL: An Interactive Framework for Reinforcement Learning for Digital Chemistry
C. Beeler, S.G. Subramanian, K. Sprague, N. Chatti, C. Bellinger, M. Shahen, N. Paquin, M. Baula, A. Dawit, Z. Yang, X. Li, M. Crowley, I. Tamblyn Digital Discovery, Advance Article (2024) Training reinforcement learning (RL) agents to perform chemical synthesis and material discovery in real laboratory settings is impractical due to the data-intensive nature of RL, the cost of physical experiments, and safety concerns. In this work, we introduce ChemGymRL, an open-source framework that provides multiple customizable virtual laboratory benches where RL agents can safely learn to perform chemistry tasks. Built on the standard Gymnasium API for broad compatibility with existing RL libraries, ChemGymRL implements well-established chemical reactions as interconnected environments that simulate realistic laboratory workflows. We train and evaluate standard RL algorithms across these virtual benches, systematically assessing their performance on chemical discovery and synthesis tasks. By providing the research community with a standardized, reusable platform, ChemGymRL bridges the gap between theoretical RL research and practical chemical discovery, enabling the development and benchmarking of RL algorithms specifically designed for chemistry applications. |
