SeisReconNO: Leveraging a U-Net-Enhanced Fourier neural operator for 3D seismic reconstruction

Missing traces in 3D seismic data are a recurring challenge caused by receiver malfunctions, acquisition limitations, and geological or environmental constraints. These gaps hinder accurate interpretation and further processing. Although numerous model-driven approaches have been developed in recent decades, they often struggle with reconstructing the data with complex geological structures and high missing ratios. To address these limitations, we proposed a U–Net-enhanced Fourier Neural Operator (UFNO), which we call SeisReconNO, a 3D seismic reconstruction framework to achieve a mesh-invariant seismic reconstruction across different missing scenarios. The SeisReconNO model leverages both spectral and spatial representations to learn a generalized reconstruction operator. We train the model on field 3D seismic cubes featuring three key missing-data patterns: random, trace-wise, and regular. Experimental results demonstrate the superior reconstruction capability of SeisReconNO across varying missing ratios. Moreover, the model exhibits strong generalization to unseen data with different resolutions, confirming its potential as a robust and adaptable tool for seismic data enhancement in real-world applications. Codes related to this paper are fully open-sourced via https://github.com/cuiyang512/SeisReconNO (We will make it public upon acceptance of this paper). • SeisReconNO reconstructs missing traces across varying 3D seismic sizes. • The model is generalized to different complex geological settings. • FNO-based operator preserves waveform continuity across varying resolutions. • Hybrid FNO–U-Net architecture recovers complex structures under high data loss. • SeisReconNO shows superior performance over U-Net in 3D seismic recovery.