Test-Time Forward Model Adaptation for Seismic Deconvolution

Seismic deconvolution is essential for extracting layer information from noisy seismic data, but it is an ill-posed problem with nonunique solutions. Inspired by classical optimization approaches, model-based deep learning architectures, such as loop unrolling (LU) methods, unfold the optimization process into iterative steps and learn gradient updates from data. These architectures rely on well-defined forward models, but in real seismic deconvolution scenarios, these models are often inaccurate or unknown. Previous approaches have addressed model uncertainty by training robust networks, either passively or actively. However, these methods require a large number of adversarial examples and diverse data structures, often necessitating retraining for unseen forward model structures, which is resource-intensive. In contrast, we propose a more efficient test-time adaptation (TTA) method for the LU architecture, which refines the forward model during inference. This approach incorporates physical principles into the reconstruction process, enabling higher quality results without the need for costly retraining.