Deep Algorithm Unrolling for Seismic Migration

Seismic migration plays a crucial role in imaging Earth’s subsurface, but traditional approaches often rely on iterative optimization methods that require manual tuning and are computationally expensive. We propose the inverse Radon transform Network (iRADNet), a deep learning-based framework that unrolls the Iterative Shrinkage Thresholding Algorithm (ISTA) into a trainable neural network, enabling automatic learning of parameters for solving the inverse Radon transform. Empirical results demonstrate the effectiveness of iRADNet in reconstructing noisy seismic data, showing enhanced accuracy and computational efficiency. Furthermore, iRADNet’s adaptable architecture allows integration with physics-informed models, opening possibilities for hybrid learning. These findings highlight the potential of deep learning-driven seismic migration for improving subsurface imaging and geophysical data interpretation. Our code and data are publicly available at https://github.com/URseismology/iRADNet.