Transformer-based wavefront sensing for atmospheric turbulence aberration correction

We propose a novel transformer-based wavefront sensing method, to the best of our knowledge, that employs a cross-task pretraining strategy to establish strong global dependencies. Compared to the CNN-based approach, this method significantly improves the aberration estimation accuracy, reducing test set loss by 70.5% and RMS by 45.7%. Notably, the attention maps of different Zernike output terms in this method exhibit remarkable consistency with the PSFs corresponding to individual aberrations. The results demonstrate that the method effectively decouples individual aberrations via the self-attention mechanism, capturing complex physical relationships and enhancing model interpretability, thus highlighting its potential as a unified methodology for advancing wavefront sensing.