Abstract Physics‐informed neural networks (PINNs) integrate physical constraints with neural architectures and leverage their nonlinear fitting capabilities to solve complex inverse problems. Tomography serves as a classic example, aiming to reconstruct subsurface velocity models to improve seismic exploration. However, the application of PINNs with multilayer perceptron architectures to such challenging inverse problems often yields suboptimal solutions. Our analysis indicates that this limitation arises from treating parameters as independent and identically distributed random variables, thereby neglecting crucial spatial correlations. To address this issue, we introduce a s pace c orrelation constrained PINN (SC‐PINN), which incorporates a spatial structure transfer network to capture essential spatial relationships during parameter estimation. SC‐PINN employs a convolutional neural network to extract spatial features from observational data, thereby better accommodating abrupt travel time changes caused by velocity discontinuities. Experiments on synthetic data sets demonstrate that SC‐PINN consistently outperforms conventional tomography method and standard PINN across a range of initial model conditions. Moreover, field experiments in the Viking Graben validate the effectiveness of SC‐PINN, delivering accurate velocity inversion results even when initialized with a simple linear model. These improvements yield clearer seabed imaging and enhanced fault characterization, underscoring the practical value of SC‐PINN for addressing complex inverse problems.
Wang et al. (Thu,) studied this question.