Calibration of wavefront aberrations using supervised learning and deep learning for a Shack–Hartmann wavefront sensor

We evaluate wavefront fitting methods in single-sample and multisample scenarios, focusing on least squares (LSQ), gradient descent, and dynamic hypernetwork approaches. LSQ is a well-established method, whereas gradient descent offers a low-cost, adaptable solution. By contrast, the dynamic hypernetwork achieves the highest precision, with a mean absolute error of 2.9559×10−5λ, significantly lower than the 2.1909×10−4λ obtained by LSQ. In addition, it reduces computation time to 0.0273 ms per sample, outperforming LSQ’s 0.4250 ms, and exhibits superior robustness in the presence of outliers. On the full testing dataset, the dynamic hypernetwork maintains consistent performance, demonstrating its ability to generalize effectively under varying levels of atmospheric turbulence. These results demonstrate that AI-based methods—particularly dynamic hypernetworks—not only address LSQ’s limitations but also enable fast and accurate wavefront fitting, making them promising tools for real-time optical applications.