Reward-Driven Optimization Framework for Physics-Informed and Explainable Image Analysis

Automated scientific imaging increasingly requires robust and explainable workflows capable of adapting to dynamic experimental conditions. We introduce a reward-driven optimization framework that formulates image analysis as a decision-making process guided by physics-informed reward functions. These rewards quantify alignment between analysis outcomes and domain objectives such as structural continuity, symmetry, or physical plausibility, enabling unsupervised, interpretable optimization of complex workflows. The framework couples classical operations with machine-learning models through Bayesian optimization to tune hyperparameters and descriptors based on reward feedback, ensuring that extracted features represent genuine physical phenomena rather than artifacts. Applied to scanning transmission electron microscopy (STEM), the approach achieves real-time segmentation and feature extraction in ion-irradiated and ​ thin films and enhances foundational models such as SAM for streaming analysis. Beyond microscopy, this methodology provides a general strategy for autonomous, physics-guided analysis across imaging and hyperspectral data, bridging explainable AI with experimental decision-making.