Work

A cold-start self-evolving multi-agent intelligence for closed-loop prediction from materials synthesis to performance

Year: 2026
Type: preprint
Abstract: Large language models (LLMs) are increasingly used to propose materials and plan experiments1–3, but whether they can predict previously unreported syntheses and the structures and performance of the resulting products before experimental validation remains unclear. Here we test this capability using a multi-agent framework of general-purpose, unfine-tuned LLMs that enables cross-disciplinary collaboration, with an AI scientist coordinating chemistry, materials science and electrochemistry agents to performance closed-loop predictions of the reaction process, crystal phase and battery performance of TiO2 anodes synthesized by a previously unreported rapid Joule-heating route. The agents predicted 64 conditions blind in eight rounds and, after each round, rewrote natural-language Skills from the released measurements. A task-matched team predicted the reaction process and phase in 91.7% and 82.8% accuracy, with errors concentrated near sharp transitions, whereas quantitative predictions of rate-capability and long-cycle performance achieved accuracies of 59.2% (72.3% for anatase-rich products) and 68.1%, within ±30% of experimental values. Revised Skills raised accuracy on identical conditions in most matched tests, and rewriting Skills without new measurements did not. A complete prediction chain cost ~US$0.27 per condition. These results indicate that LLM predictions could replace exploratory trials under conditions far away from sharp transitions, pointing to a prediction-first research paradigm.
Language: English
Open Access status: green
OpenAlex ID: W7234788759
Cites: 4
Cited by:
Related to: 10