Multichannel Wavefield Reconstruction With Physics‐Informed Neural Networks and Transfer Learning

ABSTRACT Multi‐component seismic data contain rich information essential for accurate subsurface imaging, but they are often sparse due to acquisition limitations and include noise. Robust interpolation techniques are therefore crucial to reconstruct missing traces and preserve wavefield integrity for reliable analysis and inversion. Thus, we propose a novel approach for interpolating multi‐component seismic measurements using a slope‐assisted physics‐informed neural network and compare it against a state‐of‐the‐art slope‐regularized sparsity‐promoting inversion algorithm. The performance of the two algorithms is assessed in the process of interpolating regularly subsampled shot gathers of the Mississippi streamer dataset with the proposed physics‐informed neural network framework, delivering slightly better results than those of the sparsity‐promoting inversion. An efficient transfer learning procedure that relies on fine‐tuning the network whilst interpolating consecutive shot gathers is employed to enable fast reconstruction of multiple shots, avoiding the need to fully retrain the network for each shot gather. The transfer learning approach allows the reconstruction across the entire dataset with an accuracy comparable to that of a fully trained physics‐informed neural network on each shot gather across the majority of the dataset. The novel approach yields substantial computational cost savings, enabling the reconstruction of a single shot at only 0.76% of the cost required by a physics‐informed neural network that would be retrained for each shot gather.