LLM-Guided Open Hypothesis Learning from Autonomous Scanning Probe Microscopy Experiments

Researchers have developed an open hypothesis-learning framework that combines symbolic regression with large language models to autonomously discover physical laws from scanning probe microscopy experiments. Rather than optimizing within predefined objectives, the system generates and evaluates candidate physical models directly from experimental data, demonstrating success in characterizing ferroelectric domain switching behavior.

This research represents a meaningful advancement in autonomous scientific discovery by shifting microscopy workflows from constrained optimization toward genuine hypothesis generation. Traditional autonomous experimentation systems operate within predetermined parameter spaces, efficiently refining measurements around fixed objectives. This work bridges symbolic regression—a technique for discovering mathematical relationships from data—with LLM-based physical reasoning, enabling the system to propose novel analytical models and evaluate their plausibility against known physics principles.

The approach addresses a fundamental limitation in materials science: the gap between raw experimental measurements and interpretable physical understanding. By processing sparse piezoresponse force microscopy data through symbolic regression, the framework generates candidate voltage-time relationships that progressively converge toward kinetically accurate domain-wall motion models. The LLM evaluator filters these candidates by assessing dimensional consistency, physical mechanisms, and scaling behavior—effectively automating what would otherwise require expert intuition.

For the materials science and nanotechnology sectors, this framework reduces the researcher-in-the-loop burden while maintaining physical rigor. Autonomous systems that can propose testable hypotheses accelerate materials discovery pipelines and lower barriers for scientists without deep domain expertise. The integration of symbolic reasoning with neural evaluation demonstrates how different AI paradigms complement each other in scientific contexts.

Looking forward, the critical question involves scaling: whether this approach generalizes across diverse material systems, experimental modalities beyond microscopy, and more complex multi-variable relationships. Validation against ground-truth physics and comparison with human expert hypotheses remains essential. Industrial adoption likely depends on achieving reliable autonomous discovery for high-value materials problems where hypothesis generation currently represents a bottleneck.

• →Open hypothesis learning combines symbolic regression with LLM evaluation to autonomously discover physical laws rather than optimize within fixed parameters
• →System demonstrated convergence toward physically consistent voltage-time growth laws from minimal initial piezoresponse force microscopy measurements
• →Framework extends autonomous experimentation from closed-loop optimization to genuine scientific hypothesis generation
• →Integration of symbolic reasoning and neural evaluation creates complementary strengths for materials discovery workflows
• →Scalability and cross-domain applicability remain key questions for broader adoption in materials science