Seismic first-arrival traveltime simulation based on reciprocity-constrained PINN

Simulating seismic first-arrival traveltime plays a crucial role in seismic tomography. First-arrival traveltime simulation usually relies on solving the eikonal equation. The accuracy of conventional numerical solvers is limited to a finite-difference approximation. In recent years, physics-informed neural networks (PINNs) have been applied to achieve this task. However, traditional PINNs encounter challenges in accurately solving the eikonal equation, especially in cases where the model exhibits directional scaling differences. These challenges result in substantial traveltime prediction errors when the traveling distance is long. To improve the accuracy of PINN in traveltime prediction, we incorporate the reciprocity principle as a constraint into the PINN training framework. Based on the reciprocity principle, which states that the traveltime between two points remains invariant when their roles as source and receiver are exchanged, we propose to apply this principle to multiple source–receiver pairs in PINN-based traveltime prediction. Furthermore, a dynamic weighting mechanism is proposed to balance the contributions of the eikonal equation loss and the reciprocity-constrained loss during the training process. This adaptive weighting evolves dynamically with the training epochs, enhancing the convergence of the training process. Experiments conducted on a simple lens velocity model, the Overthrust velocity model, and a 3D velocity model demonstrate that the introduction of the reciprocity-constrained PINN significantly improves the accuracy of traveltime predictions.