Autonomous Scientific Discovery via Exploration Schema Libraries: Replicating Scientists' Problem-Space Metacognition

Current Artificial Intelligence for Science (AI4S) frameworks excel at pattern recognition within massive datasets but fundamentally lack causal reasoning and the capacity for autonomous theoretical innovation. To address this limitation, we introduce a novel paradigm that shifts the focus from learning scientific data to learning the metacognitive exploration behaviors of historical scientists. We propose modeling a scientist's cognitive journey—such as proposing heuristic conjectures, navigating complex problem spaces, and inventing paradigm-shifting representational tools (e.g., Feynman diagrams)—into a formalized, reusable "Exploration Schema Library". By designing a dual-loop mechanism powered by Inverse Reinforcement Learning, AI can harvest these high-level strategic patterns from scientific history and structurally project them onto modern unresolved theoretical frameworks. This approach offers a distinct, cognitively grounded pathway toward achieving autonomous causal discovery in next-generation artificial intelligence.