Fourier-DeepONet: Fourier-enhanced deep operator networks for full waveform inversion with improved accuracy, generalizability, and robustness

Min Zhu, Shihang Feng, Youzuo Lin, Lu Lu

Abstract

Full Waveform Inversion (FWI) is a critical yet computationally intensive task in geophysics. While deep learning models offer speed, they often lack robustness to varying source parameters and noise. We present Fourier-DeepONet, a novel architecture that treats FWI as an operator learning task. By incorporating source frequency and location into a Branch-Trunk network and utilizing a Fourier-enhanced decoder, our model achieves superior generalization and resilience to data imperfections compared to traditional CNN-based approaches.

1. Introduction

Full Waveform Inversion (FWI) is the "gold standard" for imaging the Earth's subsurface. By analyzing how seismic waves travel through the ground, geophysicists can create high-resolution maps of geological structures, essential for resource exploration and monitoring the Earth's crust. However, FWI is computationally expensive and mathematically "ill-posed."

While data-driven deep learning models (like CNNs) have accelerated FWI, they suffer from a major flaw: brittleness. If a model is trained on seismic data generated by a specific source (e.g., a 15 Hz vibration at a specific spot) and then used on field data with a different frequency or location, the accuracy plummets. This lack of generalization and sensitivity to noise has hindered the practical adoption of AI in geophysical surveys.

The relationship between velocity maps and seismic data across different sources and data qualities, illustrating the FWI challenge.

2. Methods

The authors introduce Fourier-DeepONet, which builds upon the Deep Operator Network (DeepONet) architecture. A standard DeepONet consists of two sub-networks: a Branch Net that processes the input seismic waveform data, and a Trunk Net that processes domain coordinates or parameters.

Instead of just using spatial coordinates in the trunk net, the authors feed in source parameters (frequency and location). This allows the model to learn the underlying operator of the wave equation rather than just a static mapping between images. To improve output quality, the standard DeepONet decoder is replaced with a Fourier-enhanced decoder (utilizing U-FNO, a combination of Fourier Neural Operators and U-Nets). This allows the model to capture both global patterns and fine-grained geological details more effectively than a "vanilla" DeepONet.

3. Datasets and Benchmarking

The study utilizes the well-known OpenFWI benchmark, covering a range of geological complexities. Crucially, the authors introduced three new benchmark datasets to test generalization: FWI-F (variable frequencies), FWI-L (variable locations), and FWI-FL (both).

Examples of velocity maps and corresponding seismic data from the OpenFWI benchmark datasets, showing varying geological styles.

The primary subsets used include:

  • FlatVel-B (FVB): Simple horizontal layers.
  • CurveVel-A (CVA): Complex curved layers.
  • CurveFault-A (CFA): Curved layers with discontinuities (faults).
  • Style-A (STA): Highly complex, realistic geological structures.

4. Key Results and Discussion

  • Operator Learning vs. Image Mapping: By explicitly including source parameters as inputs, the model learns the "physics" of how different sources affect the resulting data. While baseline models (InversionNet and VelocityGAN) work well at exactly 15 Hz, their error increases by nearly 10x if the frequency shifts even slightly. Fourier-DeepONet remains stable across the entire 5–25 Hz range.
  • Accuracy: On the FWI-L dataset (variable locations), Fourier-DeepONet achieved an $L^2$ relative error of ~0.07–0.20, while baseline models hovered around 0.30–0.58.
  • Robustness: Because the Fourier layers operate in the frequency domain, they naturally filter out high-frequency noise. Even with 10% Gaussian noise added to the input, Fourier-DeepONet's accuracy was almost unaffected. Furthermore, the model provided satisfactory predictions even when up to half of the seismic receivers (35 out of 70) were missing data.

5. Conclusion

Fourier-DeepONet represents a significant step toward flexible and robust AI-driven geophysical surveys. While the current study focuses on the Ricker wavelet, future work will explore more diverse source shapes and receiver adaptability. Integrating governing wave equations directly into the training process (Physics-Informed Machine Learning) remains a promising path to reduce the need for massive labeled datasets.

Summary updated

BibTeX
@misc{zhu2023fourier,
  author = {Min Zhu and Shihang Feng and Youzuo Lin and Lu Lu},
  title = {{Fourier-DeepONet: Fourier-enhanced deep operator networks for full waveform inversion with improved accuracy, generalizability, and robustness}},
  year = {2023},
  eprint = {2305.17289},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2305.17289},
}
Figure