Reproduce “MACS: Multi-Agent Reinforcement Learning for Optimization of Crystal Structures” on OpenResearch

About this paper

Elena Zamaraeva, Christopher M. Collins, George R. Darling, Matthew S. Dyer, Bei Peng, Rahul Savani, Dmytro Antypov, Vladimir V. Gusev and 3 more

A multi-agent reinforcement learning approach named MACS optimizes periodic crystal structures by modeling each atom as an independent agent in a partially observable Markov game. This method achieves geometry optimization that is 34% faster and requires 28% fewer energy calculations than established baselines, while maintaining a low 0.36% failure rate and demonstrating zero-shot transferability to new chemical compositions.

Key insights

• Formulating crystal structure optimization as a MARL problem, where atoms are agents, effectively captures the localized yet interdependent nature of atomic interactions in periodic systems.
• A comprehensive observation space incorporating local atomic features (e.g., gradient, history, local environment) and an action space with an adaptive gradient-dependent scaling factor are crucial for stable and sample-efficient learning.
• The learned decentralized policy demonstrates strong zero-shot transferability to unseen chemical compositions and scalability to larger structures, highlighting its potential as a general-purpose optimizer.

Summary by alphaXiv.