Autonomous mechanism discovery from minimal experiments

<jats:p>Reliable experimental data in the physical sciences are costly and slow to obtain, limiting data-driven artificial intelligence (AI) approaches that depend on large training datasets and making evidence efficiency essential. Mechanism-driven discovery instead extracts physical rules from sparse observations, but large language models (LLM)-generated hypotheses often lack quantitative physical constraints, remaining plausible yet unverifiable. Here we introduce MiMEDAL (Minimal-experiment Mechanism-Evolving Discovery with Adaptive Learning), a nested AI framework integrating adaptive experimentation, symbolic regression, LLM reasoning and first-principles falsification. Domain-specific digital models extract expressions from minimal data to constrain LLM-generated hypotheses, which simulations iteratively test and revise until a consistent mechanism emerges. We deployed MiMEDAL on an autonomous robotic platform to investigate fluorescence in Schiff-base covalent organic frameworks (COFs). Without prior task-specific data, the system halted autonomously after 25 experiments, identifying a governing principle: emission requires exciton localisation on the aldehyde fragment, whereas leakage to the amine causes quenching. Distilled into the parameter-free descriptor η, this mechanism achieves 96.1% accuracy across 103 unseen COFs. Informed by η, the platform identified and synthesised BT-3Phethene, a conjugated COF with a 61% solid-state photoluminescence quantum yield, surpassing all reported conjugated COFs. MiMEDAL thus provides a transferable route from statistical prediction to physics-grounded mechanism discovery with minimal experimental data.</jats:p>