Bayesian Full-waveform Monitoring of CO2 Storage with Fluid-flow Priors via Generative Modeling
Abstract
Quantitative monitoring of subsurface changes is essential for ensuring the safety of geological CO2 sequestration. Full-waveform monitoring (FWM) can resolve these changes at high spatial resolution, but conventional deterministic inversion lacks uncertainty quantification and incorporates only limited prior information. Deterministic approaches can also yield unreliable results with sparse and noisy seismic data. To address these limitations, we develop a Bayesian FWM framework that combines reservoir flow physics with generative prior modeling. Prior CO2 saturation realizations are constructed by performing multiphase flow simulations on prior geological realizations. Seismic velocity is related to saturation through rock physics modeling. A variational autoencoder (VAE) trained on the priors maps high-dimensional CO2 saturation fields onto a low-dimensional, approximately Gaussian latent space, enabling efficient Bayesian inference while retaining the key geometrical structure of the CO2 plume. Hamiltonian Monte Carlo (HMC) is used to infer CO2 saturation changes from time-lapse seismic data and to quantify associated uncertainties. Numerical results show that this approach improves inversion stability and accuracy under extremely sparse and noisy acquisition, whereas deterministic methods become unreliable. Statistical seismic monitoring provides posterior uncertainty estimates that identify where additional measurements would most reduce ambiguity and mitigate errors arising from biased rock physics parameters. The framework combines reservoir physics, generative priors, and Bayesian inference to provide uncertainty quantification for time-lapse monitoring of CO2 storage and other subsurface processes.
Summary
This paper addresses the challenge of quantitative monitoring of subsurface changes during geological CO2 sequestration, which is crucial for ensuring safety and efficiency. Traditional deterministic full-waveform monitoring (FWM) methods lack uncertainty quantification and struggle with sparse and noisy seismic data, leading to unreliable results. To overcome these limitations, the authors propose a Bayesian FWM framework that integrates reservoir flow physics with generative prior modeling using a variational autoencoder (VAE). They generate prior CO2 saturation realizations through multiphase flow simulations on prior geological realizations and link seismic velocity to saturation via rock physics modeling. The VAE maps these high-dimensional saturation fields into a low-dimensional, approximately Gaussian latent space, enabling efficient Bayesian inference using Hamiltonian Monte Carlo (HMC). The key findings demonstrate that this approach improves inversion stability and accuracy, even with sparse and noisy data, where deterministic methods fail. The Bayesian framework provides posterior uncertainty estimates, allowing for the identification of locations where additional measurements would be most beneficial. Numerical experiments confirm that the VAE effectively captures the complex, non-linear spatial structures of CO2 saturation, and HMC sampling in the latent space enables efficient posterior exploration. The paper highlights the importance of incorporating physics-consistent priors for robust and reliable CO2 monitoring, especially under challenging field conditions. This framework combining reservoir physics, generative priors, and Bayesian inference provides uncertainty quantification for time-lapse monitoring of CO2 storage and other subsurface processes.
Key Insights
- •A novel Bayesian FWM framework is introduced, integrating geostatistics, fluid-flow simulations, rock physics, and deep generative modeling for improved CO2 storage monitoring.
- •A VAE is used to learn a compact, low-dimensional representation of CO2 saturation fields, enabling efficient HMC sampling while preserving the key geometrical structures of the plume. Dimensionality reduction by a factor of 40 was achieved (2685 parameters reduced to 64 latent dimensions).
- •The use of HMC in the latent space allows for gradient-based posterior exploration, improving sampling efficiency compared to traditional MCMC methods. HMC scaling is O(n^(5/4)) versus O(n^2) for traditional Metropolis-Hastings algorithms.
- •Numerical results show that the Bayesian approach improves inversion stability and accuracy under extremely sparse and noisy acquisition, where deterministic methods become unreliable.
- •The framework provides posterior uncertainty estimates, identifying areas where additional measurements would most reduce ambiguity and mitigate errors from biased rock physics parameters.
- •Latent-space interpolation experiments demonstrate that the VAE learns a smooth and continuous latent space, as quantified by the Structural Similarity Index Measure (SSIM).
- •The study evaluates the sensitivity of the framework to survey geometry, data noise, and modeling errors, demonstrating its robustness under challenging conditions.
Practical Implications
- •This research has direct applications in the real-world monitoring of geological CO2 sequestration, ensuring the long-term safety and efficiency of carbon storage operations.
- •Reservoir engineers and geophysicists would benefit from this research, as it provides a more reliable and accurate method for monitoring CO2 plume migration and assessing storage integrity.
- •Practitioners can use the proposed framework to optimize seismic survey design by identifying the most effective source locations and well placements, based on uncertainty quantification.
- •The framework opens up future research directions, including the incorporation of stochastic rock physics models, the investigation of alternative deep learning architectures, and the application to other subsurface monitoring problems such as hydrocarbon production and aquifer recharge.
- •The work provides a methodology for integrating diverse data sources within a Bayesian workflow, offering a probabilistic approach for monitoring CO2 storage and other subsurface processes that can be adapted for other relevant problems.