Unified, Practical, and White-box Seismic Tomography with Automatic Differentiation

ICLR 2026 Conference Submission16972 Authors

Keywords: seismology, automatic differentiation, inverse problem, seismic tomography

TL;DR: We present a unified, practical, white-box seismic imaging framework based on automatic differentiation, requiring no gradient derivations or neural network training.

Abstract:

Seismic tomography methods are complex, diverse, and incompatible with each other.
Traditional adjoint approaches are case-specific, requiring challenging analytical derivations for each set of parameters, waves, and loss functions.
Approximating wave equation propagation with neural networks (NNs) remains impractical, since finite training datasets cannot cover all seismic parameters for the infinite number of possible geologic models.
In this paper, we propose a unified seismic tomography framework with automatic differentiation (AD) for gradient computation, avoiding analytical derivations and NN training.
Our framework is designed for generalized misfit functionals and wave equations, supporting broader applications than previous AD-based studies.
Our method is fully white-box, and AD gradients are proven to be equivalent to adjoint gradients theoretically and numerically.
To show its generality, we performed ten cross-scenario tests across domains (time/frequency), waves (acoustic/SH/P-SV/visco-acoustic/visco-elastic), and losses (waveform/travel time/amplitude).
We also evaluated our method on the OpenFWI benchmark dataset to compare with NN methods.
Practicality was further demonstrated by a checkerboard test in the Nankai subduction zone, which is challenging for NN methods due to the lack of suitable training datasets.
Our method avoids laborious derivation and implementation of adjoint methods, with only modest computational overhead (1.3–1.8 $\times$ slower and 1.3–2.0 $\times$ more memory without mini-batching or checkpointing in our tests), which can be further reduced with these standard optimizations.
We open-sourced a PyTorch-based platform with various extensible wave simulations and imaging methods, facilitating further developments.
Our work illustrates AD's unifying capability in inverse problems, suggesting broader applications in allied scientific computing fields.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 16972