A ChemTwin System Enabling Collective Intelligence for Materials Discovery

Artificial intelligence (AI)-driven autonomous laboratories are transforming research paradigms and accelerating materials discovery 1-6. However, they still face several critical challenges, including the geographical distribution of experimental platforms with limited resource sharing, a lack of robust data privacy frameworks, and barriers to effective global collaboration for training and prediction. Here, we present ChemTwin, a system that coordinates geographically separated autonomous laboratories for collaborative materials discovery. We demonstrate its performance through three representative experiments that assess data consistency, timeliness, and multi-laboratory collaboration. By integrating a federated learning (FL) module, ChemTwin allocates experimental tasks across laboratories to increase platform utilization while preserving data privacy without any raw data exchange. Furthermore, federated learning-enhanced Bayesian optimization (FBO) achieves more than a twofold reduction in the number of experimental iterations compared with single-node Bayesian optimization (SBO), as illustrated by the optimization of a bismuth-based metal-organic framework (MOF) formulation for peroxymonosulfate (PMS) activation in organic pollutant degradation. This study builds a data privacy-preserving framework to share experimental resources across autonomous laboratories for scalable AI-driven materials research.