Self-Flow-Matching assisted Full Waveform Inversion

Xinquan Huang, Paris Perdikaris

Abstract

Full-waveform inversion (FWI) is a high-resolution seismic imaging method that estimates subsurface velocity by matching simulated and recorded waveforms. However, FWI is highly nonlinear, prone to cycle skipping, and sensitive to noise, particularly when low frequencies are missing or the initial model is poor, leading to failures under imperfect acquisition. Diffusion-regularized FWI introduces generative priors to encourage geologically realistic models, but these priors typically require costly offline pretraining and can deteriorate under distribution shift. Moreover, they assume Gaussian initialization and a fixed noise schedule, in which it is unclear how to map a deterministic FWI iterate and its starting model to a well-defined diffusion time or noise level. To address these limitations, we introduce Self-Flow-Matching assisted Full-Waveform Inversion (SFM-FWI), a physics-driven framework that eliminates the need for large-scale offline pretraining while avoiding the noise-level alignment ambiguity. SFM-FWI leverages flow matching to learn a transport field without assuming Gaussian initialization or a predefined noise schedule, so the initial model can be used directly as the starting point of the dynamics. Our approach trains a single flow network online using the governing physics and observed data. At each outer iteration, we build an interpolated model and update the flow by backpropagating the FWI data misfit, providing self-supervision without external training pairs. Experiments on challenging synthetic benchmarks show that SFM-FWI delivers more accurate reconstructions, greater noise robustness, and more stable convergence than standard FWI and pretraining-free regularization methods.

Introduction

Seismic imaging seeks to reconstruct the Earth's subsurface velocity structure by analyzing recorded sound waves. The gold standard for this task is Full Waveform Inversion (FWI). Despite its high-resolution potential, FWI is notoriously sensitive to the initial model and prone to "cycle skipping"—a phenomenon where the optimization gets trapped in local minima due to the highly non-linear relationship between velocity models and recorded waveforms.

Recent attempts to stabilize FWI have utilized generative priors, such as Diffusion Models, to guide the inversion toward geologically plausible solutions. However, these methods often require massive external datasets for pre-training and struggle with the "alignment problem": mapping a deterministic physical starting model to a specific noise level in a diffusion schedule is mathematically ambiguous. This paper introduces Self-Flow-Matching (SFM-FWI), a framework that bypasses these issues by learning a physics-driven transport field online.

Methodology: Self-Flow-Matching

SFM-FWI replaces the denoising process of diffusion models with Flow Matching. Instead of removing noise, the model learns a vector field that acts as a "GPS," calculating the most efficient path to transport a blurry initial guess toward a crisp, data-consistent target.

The SFM-FWI framework consists of an outer sampling loop and inner FWI steps. The flow network is trained online to map the initial model $m_0$ to the target $m_1$ by minimizing the data misfit through backpropagation.

Key Innovations

  • Online Training: The neural network (a U-Net) is trained during the inversion process. It requires zero offline pre-training on external datasets, relying solely on the physics of the current experiment.
  • Self-Supervision: At each step, the framework creates an interpolated version of the current model. The flow network is then updated by backpropagating the "data misfit"—the difference between simulated and recorded waveforms.
  • Elimination of Gaussian Assumptions: Because it uses flow matching, the system starts directly from the best available initial guess rather than random noise, resolving the alignment ambiguity inherent in diffusion-based FWI.

Experimental Results

The authors evaluated SFM-FWI on three classic synthetic benchmarks: Marmousi, SEG/EAGE Overthrust, and the high-contrast Salt Body Model.

Performance and Robustness

SFM-FWI consistently outperformed standard FWI and other pre-training-free methods like Deep Image Prior (DIP):

  • Accuracy: On the Marmousi benchmark, SFM-FWI achieved a Relative L2 error of 0.0130, significantly lower than standard FWI (0.0213). The Structural Similarity Index (SSIM) reached 0.907, compared to 0.853 for standard FWI.
  • Noise Resilience: Even with a signal-to-noise ratio (SNR) as low as 3.5 dB, SFM-FWI produced clear images, whereas standard methods were overwhelmed by artifacts.
  • Convergence Stability: In tests with poor initial guesses (e.g., a simple linear velocity ramp), standard FWI failed completely, while SFM-FWI maintained stable, accurate reconstructions.

Emergent Behaviors

The framework naturally exhibits automatic multi-scale inversion. It resolves large, low-frequency structures first before filling in fine details, a behavior that typically requires manual engineering in traditional geophysics workflows. Additionally, it demonstrates a "layer stripping" effect, intuitively recovering shallow structures before deeper ones.

Conclusion and Future Work

SFM-FWI provides a robust, physics-driven alternative to traditional and diffusion-based seismic inversion. By training online and utilizing flow matching, it avoids the data hunger and alignment issues of previous AI approaches. While currently demonstrated on 2D acoustic models with constant density, future work aims to extend this framework to 3D elastic physics and real-world field data, where complex attenuation and source uncertainties will provide further tests for the flow-matching paradigm.

Summary updated

Abstract

Full-waveform inversion (FWI) is a high-resolution seismic imaging method that estimates subsurface velocity by matching simulated and recorded waveforms. However, FWI is highly nonlinear, prone to cycle skipping, and sensitive to noise, particularly when low frequencies are missing or the initial model is poor, leading to failures under imperfect acquisition. Diffusion-regularized FWI introduces generative priors to encourage geologically realistic models, but these priors typically require costly offline pretraining and can deteriorate under distribution shift. Moreover, they assume Gaussian initialization and a fixed noise schedule, in which it is unclear how to map a deterministic FWI iterate and its starting model to a well-defined diffusion time or noise level. To address these limitations, we introduce Self-Flow-Matching assisted Full-Waveform Inversion (SFM-FWI), a physics-driven framework that eliminates the need for large-scale offline pretraining while avoiding the noise-level alignment ambiguity. SFM-FWI leverages flow matching to learn a transport field without assuming Gaussian initialization or a predefined noise schedule, so the initial model can be used directly as the starting point of the dynamics. Our approach trains a single flow network online using the governing physics and observed data. At each outer iteration, we build an interpolated model and update the flow by backpropagating the FWI data misfit, providing self-supervision without external training pairs. Experiments on challenging synthetic benchmarks show that SFM-FWI delivers more accurate reconstructions, greater noise robustness, and more stable convergence than standard FWI and pretraining-free regularization methods.
BibTeX
@misc{huang2026self,
  author = {Xinquan Huang and Paris Perdikaris},
  title = {{Self-Flow-Matching assisted Full Waveform Inversion}},
  year = {2026},
  eprint = {2603.13425},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.13425},
}
Figure