MANDREL: Modular Reinforcement Learning Pipelines for Material Discovery

Abstract

AI-driven Material Discovery is evolving rapidly with new approaches and pipelines for experimentation and design. However, the pipelines are often designed in isolation. It is therefore difficult to compare the performance of different pipelines or leverage the advance of one pipeline for another. We introduce a modular reinforcement learning framework for interoperable experimentation and design of tailored, novel molecular species. The framework unifies disparate approaches to RL of molecular graphs by offering a comprehensive set of modules allowing mixing and matching choices for the underlying chemical action space, molecular representations, targeted molecular properties, and common RL algorithms. Our demo showcases the framework’s capabilities applied to benchmark problems like QED and plogp, as well as the design of novel small molecule solvents for carbon capture. The framework is designed for plug-and-play experimentation and is compatible with the popular SB3 and RLLIB libraries.