Abstract:

Traveltime tomography is a key technique in seismology and seismic exploration. However, traditional traveltime tomography methods face numerous challenges, such as inaccuracies in imaging results due to spatial resolution, numerical errors, and data matching after model discretization, as well as limitations in the efficiency of multi-source simulations. To overcome these limitations, this study employs physics-informed neural networks (PINNs) to solve the eikonal equation for simulating seismic traveltimes and implementing traveltime tomography. This PINN-based tomography approach is named PINNtomo. PINNtomo integrates physical constraints (the eikonal equation) and data constraints (travel times) into the loss function, enabling rapid and accurate traveltime field simulation and velocity modeling without relying on an initial model. This paper investigates the impact of multiplicative and additive traveltime T factorization on the accuracy and efficiency of PINNtomo. The results demonstrate that as the model size increases, the additive factorization method not only significantly improves imaging efficiency (the reduction in computation time for additive PINNtomo increases from 92 s to 3 486 s compared to multiplicative PINNtomo), but also enhances accuracy and optimizes computational performance.

Key words: physics-informed neural networks')">

physics-informed neural networks, tomography, inversion, traveltime, eikonal equation