Multiscale Data-Driven Seismic Full-Waveform Inversion With Field Data Study

Shihang Feng, Youzuo Lin, Brendt WohlbergIEEE TGRS 2022 · DOI

TL;DR

This paper introduces a novel "Multiscale InversionNet" that leverages deep learning and style transfer to create realistic subsurface velocity models from seismic data. By decoupling the inversion into low- and high-resolution components and using natural images to generate diverse training data, the method significantly reduces computation time and improves accuracy compared to traditional seismic full-waveform inversion (FWI), even on real field data.

What Problem Does This Paper Address?

Seismic full-waveform inversion (FWI) is a powerful technique for imaging subsurface geological structures, crucial for applications like energy exploration and infrastructure assessment. However, traditional FWI faces significant challenges:

  • Computational Cost: FWI is computationally intensive, requiring vast resources, especially for high-resolution seismic data.
  • Non-Convexity and Local Minima: The inversion process is non-convex, meaning it can get stuck in suboptimal solutions (local minima), particularly if the initial velocity model is inaccurate or lacks low-frequency information.
  • Data Scarcity: Creating realistic training datasets for data-driven FWI is difficult and expensive, as it requires matching subsurface models with corresponding seismic measurements. This leads to poor generalization capabilities for data-driven methods when applied to new, unseen data.

This paper aims to overcome these limitations by developing a data-driven FWI approach that is computationally efficient, robust to initial model inaccuracies, and generalizes well to field data.

Methods

The proposed solution, "Multiscale InversionNet," combines several innovative techniques:

  • Data Augmentation via Style Transfer: Instead of relying solely on limited real subsurface models, the authors use a style transfer technique (inspired by artistic style transfer) to generate a large, diverse dataset of "physically practical" subsurface velocity models. Natural images (from the COCO dataset) are used as content, and a known geological model (like Marmousi) provides the style, creating realistic geological features.
  • Multiscale Inversion Strategy: The inversion process is split into two stages, mimicking multiscale approaches in traditional FWI:
  • Low-resolution InversionNet: This network first inverts for the low-frequency components of the velocity model, focusing on direct and transmission waves. This provides a smooth, initial estimate.
  • High-resolution InversionNet: This network then refines the low-resolution model by incorporating reflection waves from the data residual. It uses the output of the low-resolution network as an initial guess, significantly improving accuracy and capturing finer details.
  • Encoder-Decoder Architectures: Both networks utilize encoder-decoder architectures, a common deep learning paradigm for image-to-image translation tasks, allowing them to learn complex mappings between seismic data and subsurface models.
  • Two-Step Training for High-Resolution Network: To ensure stable training, the high-resolution network is trained in two steps: first, fixing the data-processing part and training the model-processing part and decoder; second, fixing the model-processing part and training the data-processing part and decoder.

Datasets

  • Training Data Generation:
    • Content Images: 67,000 natural images from the COCO dataset [52].
  • Style Images: The Marmousi velocity model [55] was used as a style reference.
  • Synthetic Seismic Data: Generated using the acoustic wave equation with a Ricker wavelet source.
  • Testing Data:
    • Synthetic Datasets: In-distribution test data, as well as out-of-distribution datasets based on Marmousi and Overthrust [59] velocity models.
  • Field Data: A 2D Gulf of Mexico (GOM) dataset [63, 64].

Key Insights and Ideas

  • Style Transfer for Realistic Data Generation: Using natural images with a geological style transfer is a powerful way to create diverse and physically plausible training data for seismic inversion, overcoming the scarcity of real subsurface models.
  • Decoupling Wave Phenomena: Separating the inversion into low-frequency (transmission/direct waves) and high-frequency (reflection waves) components allows for a more robust and accurate reconstruction, especially in complex geological settings.
  • Data-Driven Efficiency: Once trained, the Multiscale InversionNet can perform inversion significantly faster than traditional physics-based FWI, making it practical for large-scale applications.
  • Improved Generalization: The diverse training data generated through style transfer leads to better generalization capabilities, allowing the model to perform well on unseen synthetic and real field data.

Key Results

  • Computational Speedup: The trained Multiscale InversionNet significantly reduces computation time compared to conventional FWI.
  • Accuracy Improvement: The method yields more accurate subsurface velocity models, as evidenced by lower Mean Squared Error (MSE) and higher Structural Similarity Index (SSIM) scores in synthetic tests (e.g., MSE of 0.0212 and SSIM of 0.7820 for a test model).
  • Field Data Performance: The method demonstrates promising results on a real 2D Gulf of Mexico dataset, producing comparable or improved results to physics-based methods like Wave Equation Traveltime Tomography (WT) and conventional FWI, particularly in capturing deeper reflection events. Reverse Time Migration (RTM) was used as a quality control tool, showing more focused and continuous reflections for the proposed method.

Limitations and Future Work

  • Limitations:
    • Global Features: While the method excels at capturing local geological features, it may not fully capture global features, partly due to the simplified 1D background model used.
  • Data Consistency: The current loss function primarily focuses on the velocity model domain. Incorporating a data consistency loss (e.g., cycle-consistency loss) could further refine subsurface structures and improve accuracy.
  • Artifacts: Some artifacts and missing information were observed in inversion results for complex geological structures (e.g., Overthrust model), suggesting the need for more diverse training data.
  • Future Work:
    • Content Image Selection: More careful selection of content natural images could further enhance the diversity and quality of generated velocity models.
  • Loss Function Refinement: Exploring different loss functions, such as cycle-consistency loss, to better account for data consistency and improve the reconstruction of fine details.
  • Physics-Based Regularization: Incorporating physics-based regularization techniques, like illumination compensation, to improve inversion accuracy, especially in deeper regions.
  • Orthogonal Styles: Using multiple subsurface style models with orthogonal features to create even more representative training datasets.

Summary updated

BibTeX
@article{feng2022multiscale,
  author = {Shihang Feng and Youzuo Lin and Brendt Wohlberg},
  title = {{Multiscale Data-Driven Seismic Full-Waveform Inversion With Field Data Study}},
  journal = {IEEE TGRS},
  year = {2022},
  doi = {10.1109/tgrs.2021.3114101},
}
Figure