Learning a robust 3D seismic data reconstruction operator using UFNO

Seismic data often suffer from missing traces due to receiver malfunctions, environmental factors, or processing challenges, particularly in 3D seismic surveys. This data loss complicates further processing and interpretation. Traditional interpolation methods like linear and Fourier interpolation, Radon transform, and wavelet-based transforms perform well but struggle with complex seismic data or high proportions of missing traces. Deep learning offers a promising data-driven alternative, but conventional approaches are limited to predicting data of the same shape as the training set, restricting their applicability in real-world scenarios. A novel workflow based on the Fourier Neural Operator (FNO) addresses these challenges by learning a reconstruction operator capable of handling seismic bins with varying resolutions. This approach targets three missing trace scenarios: random inline/crossline gaps, fully random missing traces, and large gap missing sections. Metrics such as mean square error (MSE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and Fréchet distance (FD) are used to evaluate performance. Experimental results demonstrate the FNO-based method's effectiveness in reconstructing data from complex geological structures, surpassing traditional and conventional deep learning methods. The workflow represents a significant advancement for seismic interpolation in diverse and challenging field conditions.