Adaptive human-in-the-loop optimization using language-guided priors for chemical experiments

The efficiency of self-driving laboratories is often limited by the cold start problem, where optimization algorithms must relearn fundamental chemical trends already known to domain experts. While incorporating prior knowledge can significantly accelerate discovery, existing methods struggle to translate qualitative expert intuition into quantitative models and lack practical safeguards against misleading heuristics. Here, we introduce the human-in-the-loop optimization framework, which utilizes large language models to parse natural language prompts into structured prior mean functions and constraints for Bayesian optimization. To reduce bias and prior lock-in, we introduce the adaptive weight credibility detection (AWCD) mechanism, which continuously evaluates the alignment between expert hypotheses (positive trends or exclusion zones) and emerging experimental data. Functioning as a probabilistic triggering mechanism, AWCD dynamically disables the prior upon detecting sufficient contradictory evidence, so that the system can recover baseline-like data-driven optimization behavior once enough counter-evidence has been sampled under the tested conditions. Validation against the continuous Ugi multicomponent reaction and discrete P3HT-CNT composite data sets demonstrates that this approach, when enriched with accurate expert knowledge, accelerates the identification of high-performing conditions relative to unguided Bayesian optimization, while the AWCD safeguard limits the performance loss incurred when the supplied prior is misleading. Furthermore, we illustrate the practical utility of the framework via an interactive "Co-Scientist" interface, which enables iterative knowledge injection and allows domain experts to guide autonomous discovery without specialized computational expertise.