Graph-based modeling of optical system enables adaptive optics on dynamic samples with self-calibration.

Sensor less adaptive optics offers significant advantages over hardware-based wavefront sensing but faces persistent challenges: Its performance degrades when idealized models fail to capture system imperfections, it is largely restricted to spatially invariant aberrations, and it cannot accommodate dynamic biological samples due to static-object assumptions. Here we present graph-modeling and phase-diversity-based computational adaptive optics with self-calibration (GRAPHYCS), a differentiable graph-based modeling framework that addresses all three limitations. GRAPHYCS automatically self-calibrates to correct system-specific non-idealities, enables spatially variant wavefront sensing across extended fields of view by modeling local aberrations, and supports dynamic live-sample imaging where conventional computational methods fail. In simulations, GRAPHYCS achieves up to a 9-fold improvement in wavefront sensing accuracy compared to analytic phase diversity under system non-idealities. In real microscopy experiments, it consistently outperforms phase-diversity-based methods compared in this study. Furthermore, in live zebrafish brain imaging, GRAPHYCS enables simultaneous wavefront sensing and neuronal activity detection-an application beyond the reach of existing approaches without additional hardware complexity.