A new method for seismic wave travel time computation based on physics-informed generative adversarial networks

Traveltime computation is fundamental to many applications in seismology, including seismic imaging, traveltime tomography, and earthquake location. We present a novel physics-informed generative adversarial network (pi-GAN) for solving the Eikonal equation in three-dimensional velocity models. The framework consists of two adversarial networks: a generator that predicts the velocity model and a discriminator that distinguishes it from the true model. A physics-based constraint derived from the Eikonal equation is imposed during training to ensure consistency between the predicted traveltime field and the velocity model. Spatial derivatives are computed via automatic differentiation, enabling mesh-free traveltime estimation between arbitrary points. The method leverages GPU acceleration for efficient, highly parallelized computation and reduced memory usage. Benchmark tests demonstrate that pi-GAN achieves higher accuracy and lower memory consumption than existing neural solvers such as EikoNet, particularly for models with strong velocity contrasts. These results highlight the potential of pi-GAN as a general-purpose, efficient solver for traveltime computation in complex media.