Hybrid attention graph neural network for dynamic spatiotemporal wavefront prediction in adaptive optics

Abstract

Adaptive optics (AO) systems, inherently constrained by delay errors, suffer from limitations in their correction performance. The proposed predictive AO technology aims to mitigate these delays, thereby enhancing the system’s correction bandwidth. However, existing predictive algorithms primarily focus on extracting temporal features while neglecting the spatial characteristics during prediction, which compromises the generalization robustness. In this paper, we introduce a hybrid attention graph neural network (HAG-Net) for dynamic spatiotemporal wavefront prediction. HAG-Net combines temporal convolution and graph convolution to effectively capture both spatiotemporal features of wavefront data, while the integration of a dynamic graph learning and attention mechanism enhances the extraction of spatial correlations between wavefront frames, resulting in superior predictive accuracy. We compared HAG-Net with two established predictive algorithms and evaluated their performance under various atmospheric conditions. In simulations, HAG-Net reduces the mean and standard deviation of root square error (RMS) by 55.8% and 61.7%, respectively, compared to traditional AO. Experimental results further demonstrate that our method increased the maximum far-field focusing intensity by approximately 4.6 times relative to the uncorrected scenario. HAG-Net consistently outperformed other algorithms in both simulated and experimental settings, positioning it as a promising solution for addressing latency challenges in closed-loop AO systems.