EKtomo: A Physics‐Informed Neural Network Approach for Robust Crosswell Travel‐Time Tomography

ABSTRACT Seismic tomography has long been a key technique for building reliable subsurface structural models. As a critical geophysical parameter with geological implications, velocity needs to be accurately estimated, particularly in near‐surface studies where complex geological processes shape the subsurface. In seismic exploration, it is often used to provide initial velocity models for subsequent depth migration and full‐waveform inversion. To overcome the limitations of traditional tomographic methods, such as instability and mesh‐dependency, EKtomo, a physics‐informed neural network approach, offers an effective alternative for building near‐surface velocity models using first‐arrival times. By incorporating an initial velocity model directly as a network input, the method provides effective prior constraints at a low computational cost. Its core principle involves simultaneously optimizing two independent neural networks, one for predicting travel‐times and the other for velocity, by minimizing two loss functions: a physical loss constrained by the eikonal equation and a loss of data misfit from travel‐time data. Through tests on both crosswell synthetic and field data, the proposed method demonstrates excellent inversion performance for the velocity field and provides an effective alternative for constructing near‐surface velocity models from first‐arrival times. The various factors controlling the performance of EKtomo and its fundamental limitations are also discussed.