Autonomous Physics Discovery via Reward-Guided Multi-Objective Deep Learning in Ferroelectrics

Discovering structure–property relationships is central to materials science but remains hindered by the limited availability of labeled data and the destructive nature of many measurements. Here, we present a materials-focused active learning framework that autonomously uncovers physically explainable links between microstructure and functionality in complex ferroelectrics. Using multi-objective Bayesian optimization with deep kernel learning (MOBO-DKL), our approach elevates physically meaningful descriptors—such as domain-boundary proximity—alongside functional responses, enabling interpretable mapping of trade-offs on the Pareto front. Applied to fully automated piezoresponse force microscopy, MOBO-DKL achieves a ~300-fold acceleration over grid-based approaches while revealing how local domain environments govern polarization switching. By embedding human physical intuition into the learning objectives, this framework transforms autonomous experimentation from efficient measurement to genuine physics discovery. Beyond ferroelectrics, the approach offers a generalizable route for accelerating functional materials research, providing insights critical for the design of next-generation memory, electromechanical, and quantum devices.