SVGym (SciVerseGym): An Environment for Reinforcement Learning and Bayesian Optimization in Crystal Discovery

Overview

SciVerseGym addresses the fragmentation in computational materials discovery, where search algorithms—such as reinforcement learning, Bayesian optimization, and evolutionary search—often rely on bespoke, non-reproducible pipelines for structure manipulation and evaluation. By adopting the Gymnasium API, SciVerseGym provides a standardized environment that treats crystal design as a Markov decision process. This allows researchers to swap different search agents and physical evaluators without modifying the underlying infrastructure.

Instead of requiring agents to generate complete crystallographic files, which is prone to errors, SciVerseGym uses a set of bounded, chemically meaningful edit actions. These include elemental substitution, lattice perturbation, atomic displacement, vacancy creation, and atom insertion. At each step, the agent proposes an edit, the environment updates the atomistic structure, and a configurable backend (such as a machine-learned interatomic potential) evaluates the candidate. The environment then returns a standard tuple containing the new observation, a scalar reward, termination/truncation flags, and diagnostic metadata.

This framework enables fair, reproducible benchmarking across diverse algorithmic families. By separating the search policy from the physical evaluation, researchers can test how different agents perform under identical physical assumptions. Furthermore, the inclusion of diagnostic information—such as energy above the convex hull and phonon stability—in the environment's output allows for more nuanced analysis of search trajectories, moving beyond simple scalar rewards to provide deeper insights into why specific materials candidates are proposed or rejected.