Deep learning-enabled computational adaptive-optics for fast continuous zoom microscopy

Conventional optical microscopes are constrained by discrete magnification settings, limiting continuous multi-scale observation. Although liquid lenses enable continuous zoom, their practical performance is often compromised by dynamic aberrations and electrowetting-induced oscillations. Here, we propose a deep learning-enabled computational adaptive-optics (AO) framework for fast continuous zoom microscopy. Our physics-inspired network directly estimates the point spread function (PSF) from degraded images by leveraging complementary frequency-domain cues and spatial gradient information. By coupling sliding window self-attention with PSF-guided dynamic filtering, the model achieves accurate image restoration across zoom states. In addition, we introduce a degradation model that simulates electrowetting liquid-surface oscillations together with realistic imaging degradations. Experiments validate that this computational AO approach accelerates continuous zoom imaging while remaining robust to vignetting-induced non-uniform illumination, thereby enabling fast, high-quality multi-scale observation in continuous zoom microscopy.