Resolving phase ambiguity via physically optimized single-image sensorless wavefront sensing with neural networks

By offering simplicity, high efficiency, and low cost, deep learning-based wavefront sensing from a single in-focus image is a highly promising alternative for adaptive optics. However, its performance is fundamentally constrained by phase ambiguity, where identical image patterns may correspond to distinct phase aberrations. Here, we propose a physically optimized single-image sensorless (POSS) method, which integrates a lightweight ResNet-18 with a physics-guided sign correction strategy. POSS effectively resolves the sign ambiguity of even-order Zernike coefficients, achieving reconstruction accuracy comparable to high-precision dual-image phase diversity methods across a broad turbulence range (<i>D</i>/<i>r</i>₀ = 1-15). Meanwhile, without requiring additional hardware, it enables millisecond-level inference, maintains stable residual RMS under varying noise conditions, and provides improved high-frequency recovery under low-to-moderate turbulence. Furthermore, two progressive sign correction strategies are introduced to balance reconstruction accuracy and computational cost. POSS enables accurate, robust, and real-time wavefront reconstruction from a single in-focus image, providing a low-complexity and hardware-friendly solution for intelligent adaptive optics systems.