Abstract Seismic traveltime tomography is a widely used approach for imaging Earth's interior at various scales. However, constructing reliable velocity models typically requires selecting suitable inversion grids and regularization parameters, a process that is often subjective and cumbersome, and may compromise model resolution. To overcome these limitations, we combine deep neural networks with conventional traveltime tomography by using an untrained network to parameterize the velocity model, which is then employed to compute synthetic travel times. The misfit between synthetic and observed travel times is minimized by updating the network parameters, and the final tomography model is obtained from the trained network. Neural networks inherently introduce spatial correlations as regularization, eliminating the need for manual tuning. Synthetic tests demonstrate that the method achieves higher accuracy than conventional methods, recovering fine‐scale structures while maintaining the inversion stability in regions with sparse data coverage. Application to seismic data from central Chile further validates the robustness of our method, revealing a clearer subducting slab with fewer artifacts.
Yang et al. (Mon,) studied this question.