Skip to Content
MathematicsMathematics
  • Article
  • Open Access

21 April 2026

22 Pages

Residual Physics-Informed Neural Networks for Seismic Tomography with Multi-Source Prior Constraints

,
,
and
School of Sciences, Southwest Petroleum University, Chengdu 610500, China
*
Author to whom correspondence should be addressed.
This article belongs to the Section E: Applied Mathematics

Abstract

Seismic traveltime tomography is essential for constructing subsurface velocity models that underpin high-resolution imaging and inversion. Traditional ray- and eikonal-based methods are sensitive to the starting model and lack a unified, physically consistent framework to integrate seismic data with high-confidence prior information. PINN-based approaches offer flexible, grid-free inversion but often suffer from training instability and limited use of prior constraints. We propose a Multi-Source Prior-Guided Residual Physics-Informed Neural Network (MSP-ResPINNs) to address these limitations through two key technical advancements. First, MSP-ResPINNs integrates a Residual Network (ResNet) architecture with Sinusoidal Representation (SIREN) activation to replace standard MLPs, ensuring the robust capture of high-frequency velocity gradients. Second, the framework implements a unified loss function that rigorously enforces multi-source constraints, including well-logs and geological horizons. Numerical experiments demonstrate that MSP-ResPINNs accurately reconstructs sharp velocity contrasts and complex geological features compared with conventional PINN-based approaches. Among the tested variants, the multiplicative factorization consistently provides the most stable and physically consistent results, outperforming the additive factorization.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.