Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Physics-guided Generative AI Toolkit for Geophysical Monitoring

Junhuan Yang, Hanchen Wang, Yi Sheng, Youzuo Lin, Lei Yang

TL;DR

This work tackles the challenge of performing accurate Full-Waveform Inversion under data scarcity and edge-resource constraints by introducing EdGeo, a physics-guided diffusion toolkit. EdGeo uses a VM-conditional diffusion model to generate velocity maps, then applies leakage movement and distribution alignment to produce unprivileged data that reflects shallow features, followed by a physics-based forward model to create paired seismic data for end-to-end fine-tuning of pruned InversionNet models. The approach improves SSIM, MAE, and MSE across pruning ratios and achieves favorable latency on edge devices, outperforming VAE, VAE-Reg, and standard diffusion baselines, especially in representing shallow leakage features. The work highlights the practical impact of integrating physics guidance with generative AI to enable real-time, localized geophysical monitoring tasks such as CO2 leakage detection.

Abstract

Full-waveform inversion (FWI) plays a vital role in geoscience to explore the subsurface. It utilizes the seismic wave to image the subsurface velocity map. As the machine learning (ML) technique evolves, the data-driven approaches using ML for FWI tasks have emerged, offering enhanced accuracy and reduced computational cost compared to traditional physics-based methods. However, a common challenge in geoscience, the unprivileged data, severely limits ML effectiveness. The issue becomes even worse during model pruning, a step essential in geoscience due to environmental complexities. To tackle this, we introduce the EdGeo toolkit, which employs a diffusion-based model guided by physics principles to generate high-fidelity velocity maps. The toolkit uses the acoustic wave equation to generate corresponding seismic waveform data, facilitating the fine-tuning of pruned ML models. Our results demonstrate significant improvements in SSIM scores and reduction in both MAE and MSE across various pruning ratios. Notably, the ML model fine-tuned using data generated by EdGeo yields superior quality of velocity maps, especially in representing unprivileged features, outperforming other existing methods.

A Physics-guided Generative AI Toolkit for Geophysical Monitoring

TL;DR

This work tackles the challenge of performing accurate Full-Waveform Inversion under data scarcity and edge-resource constraints by introducing EdGeo, a physics-guided diffusion toolkit. EdGeo uses a VM-conditional diffusion model to generate velocity maps, then applies leakage movement and distribution alignment to produce unprivileged data that reflects shallow features, followed by a physics-based forward model to create paired seismic data for end-to-end fine-tuning of pruned InversionNet models. The approach improves SSIM, MAE, and MSE across pruning ratios and achieves favorable latency on edge devices, outperforming VAE, VAE-Reg, and standard diffusion baselines, especially in representing shallow leakage features. The work highlights the practical impact of integrating physics guidance with generative AI to enable real-time, localized geophysical monitoring tasks such as CO2 leakage detection.

Abstract

Full-waveform inversion (FWI) plays a vital role in geoscience to explore the subsurface. It utilizes the seismic wave to image the subsurface velocity map. As the machine learning (ML) technique evolves, the data-driven approaches using ML for FWI tasks have emerged, offering enhanced accuracy and reduced computational cost compared to traditional physics-based methods. However, a common challenge in geoscience, the unprivileged data, severely limits ML effectiveness. The issue becomes even worse during model pruning, a step essential in geoscience due to environmental complexities. To tackle this, we introduce the EdGeo toolkit, which employs a diffusion-based model guided by physics principles to generate high-fidelity velocity maps. The toolkit uses the acoustic wave equation to generate corresponding seismic waveform data, facilitating the fine-tuning of pruned ML models. Our results demonstrate significant improvements in SSIM scores and reduction in both MAE and MSE across various pruning ratios. Notably, the ML model fine-tuned using data generated by EdGeo yields superior quality of velocity maps, especially in representing unprivileged features, outperforming other existing methods.
Paper Structure (4 sections, 5 equations, 6 figures)

This paper contains 4 sections, 5 equations, 6 figures.

Table of Contents

  1. Introduction
  2. Framework
  3. Experiment
  4. Conclusion

Figures (6)

  • Figure 1: An example of (a) Seismic exploration (b) Seismic data and its corresponding (c) Velocity map.
  • Figure 2: Overview of the end-to-end fine-tuning framework and EdGeo toolkit
  • Figure 3: Leakage movement and Distribution alignment
  • Figure 4: Performance with 95% and 90% pruning ratio
  • Figure 5: Comparison between the EdGeo and competitors
  • ...and 1 more figures