Few-shot learning-based rapid spatiotemporal wavefront prediction with a lightweight Fourier graph neural network in adaptive optics

Adaptive optics (AO) systems hold significant application values, but are fundamentally limited by the inherent time delay challenge between wavefront sensing and correction. Prediction algorithms have emerged to infer future wavefront evolution, thereby enabling predictive AO. However, existing algorithms often face trade-offs among accuracy, data efficiency, generalization capacity, and response bandwidth. To address these limitations, here we propose, to our knowledge, a novel few-shot learning framework for spatiotemporal wavefront prediction, termed the Fourier graph neural network (FGNN). FGNN uniquely enables accurate forecasting of wavefront evolution using only few-shot wavefront frames. It also exhibits strong robustness under varying turbulence conditions, including changes in wind speed, Fried parameter, and wavefront sensor frequency. Among all comparative algorithms, FGNN possesses the lowest parameter count, the minimum computational complexity, and the shortest response time of 0.121 ms. It reduces the residual wavefront error by up to 70% and improves the peak intensity of the focused spot by ∼30 % in an experimental open-loop AO setup. By providing a data-efficient and computationally lightweight solution, FGNN serves as a powerful tool for advanced AO technology in specific scenarios, with significant potential for applications in astronomical observation and laser communications.