Large-language-model-driven adaptive search space definition for autonomous closed-loop materials exploration

Autonomous closed-loop materials exploration systems that couple machine learning with robotic experimentation, often called self-driving laboratories, are increasingly used to accelerate materials discovery. However, their end-to-end autonomy is limited by the need for human experts to specify and periodically reconfigure the search space. Herein, we propose a data-driven framework that adaptively redefines the materials search space during operation for continuous exploration. It comprises a large language model (LLM)-driven outer loop that proposes updated search spaces and an inner Bayesian optimization loop that selects candidate materials within each space. This hierarchical scheme enables efficient exploration across substantially expanded search spaces. Target properties for which the LLM possesses extensive prior knowledge exhibit the most pronounced performance gains, suggesting that foundation models can leverage learned domain knowledge to generate informative search space updates. This framework could be integrated into diverse autonomous materials exploration platforms and represents a step toward more autonomous exploration across broader materials search spaces. Self-driving labs speed materials discovery but still rely on experts to define and update the search space. Here, the authors propose a large language model-guided Bayesian optimization framework that adaptively updates the search space for more autonomous and efficient materials exploration.