Physics-guided reinforcement learning for structured illumination microscopy

Abstract Structured illumination microscopy improves fluorescence imaging by shifting fine specimen information into the observable passband, but reconstructions often deteriorate when illumination phases, fringe contrast or noise depart from calibrated conditions. Existing learning-based methods usually compensate for these imperfections only after acquisition. Here we show a physics-guided reinforcement-learning framework for structured illumination microscopy that couples a differentiable optical forward model, an encoder–decoder reconstructor and a Soft Actor–Critic controller during training. The controller adaptively perturbs illumination phase, modulation depth and pattern frequency within physical bounds, while the reconstructor is optimised with image-domain, measurement-domain and spectral constraints. On simulated BioSR data, the method improves structural fidelity and frequency recovery relative to wide-field references and learning-based baselines, and remains stable under noise, phase detuning, stripe interference and photobleaching. Experiments on fixed-cell and bead samples acquired with a digital micromirror device platform indicate transfer to hardware without experimental fine-tuning.