Physics-informed neural network for reconstruction of seismic data under compressed sensing sampling

Seismic exploration is one of the most critical methodologies and the highest-cost expenditures in the pre-exploration. The main cost of seismic exploration is acquiring seismic data, which can be significantly reduced through compressed sensing (CS) techniques. Traditional and deep learning (DL) CS methods offer unprecedented opportunities for cost optimization while maintaining data fidelity. However, CS methods rely on random acquisition, which performs poorly when the seismic data are not randomly acquired. This manuscript proposes a novel physics-informed neural network (PINN) framework for reconstructing 3D seismic data acquired via down-sampling from Ocean Bottom Seismometer (OBS) observation systems. The compressed sensing acquisition system of seismic data contains two types of sparsity: 1) 2D random missing traces, 2) Dual random missing of source lines and source points. The proposed method employed move-out (MO) transformations with multiple constant velocities to mitigate aliasing artifacts and improve reconstruction accuracy. Then, a pre-interpolation process is utilized for the MO-transformed seismic data groups. Additionally, a semblance evaluation mechanism dynamically assigns weights to each MO dataset, generating optimized, pre-interpolated seismic profiles. Finally, the PINN architecture integrates physical constraints to refine the reconstructed data. The experimental results demonstrate the superior reconstruction performance and computational efficiency of the proposed method compared with the state-of-the-art. This manuscript has proposed a novel physics-informed neural network (PINN) framework for reconstructing 3D seismic data acquired via down-sampling from the Ocean Bottom Seismometer (OBS) observation system. The compressed sensing seismic data contains two types of sparsity: 1) 2D random missing traces, 2) Dual random missing of source lines and source points. The proposed method employed move-out (MO) transformations with multiple constant velocities to mitigate aliasing artifacts and improve reconstruction accuracy. Then, a pre-interpolation process is utilized for the MO-transformed seismic data groups. Additionally, a semblance evaluation mechanism dynamically assigns weights to each MO dataset, generating optimized, pre-interpolated seismic profiles. Finally, the PINN architecture integrates physical constraints to refine the reconstructed data. The pre-interpolation model is trained on plenty of unpaired data in a semi-supervised way, while the refined model learns on the target data in a self-supervised way. • The compressed sensing seismic data contains two types of sparsity: 1) 2D random missing traces, 2) Dual random missing of source lines and source points. • The proposed method utilized move-out (MO) transformations with multiple constant velocities to address aliasing artifacts and enhance reconstruction accuracy. • A deep neural network is applied to pre-interpolate seismic data after each MO process in a semi-supervised way. • Physics-informed neural network is employed to fuse the pre-interpolated seismic data and refine the reconstructed results by physical constraints.