Diffusion-based subsurface CO$_2$ multiphysics monitoring and forecasting

TL;DR

The paper introduces a diffusion-based video generation framework to model coupled subsurface processes from CO2 injection, enabling efficient forecasting, inversion, and uncertainty quantification by learning the joint distribution $p(\mathbf{x}^{1:T}|\mathbf{y})$ over 3D multiphysics volumes. It treats permeability, CO2 saturation, velocity, density, and RTM images as a time-evolving video and uses reconstruction-guided conditional sampling to incorporate history and observations without retraining. Through experiments on the Compass synthetic benchmark, the approach demonstrates unconditional generation, accurate forecasting (avg relative L2 ~$0.024$, SSIM ~$0.889$), reliable inversion (avg relative L2 ~$0.021$, SSIM ~$0.961$), and meaningful uncertainty maps, while acknowledging substantial per-sample computational costs. The work offers a potentially transformative, end-to-end tool for CCS monitoring and decision support, with future directions including physics-informed losses, longer horizons, and more realistic field conditions.

Abstract

Carbon capture and storage (CCS) plays a crucial role in mitigating greenhouse gas emissions, particularly from industrial outputs. Using seismic monitoring can aid in an accurate and robust monitoring system to ensure the effectiveness of CCS and mitigate associated risks. However, conventional seismic wave equation-based approaches are computationally demanding, which hinders real-time applications. In addition to efficiency, forecasting and uncertainty analysis are not easy to handle using such numerical-simulation-based approaches. To this end, we propose a novel subsurface multiphysics monitoring and forecasting framework utilizing video diffusion models. This approach can generate high-quality representations of CO$2$ evolution and associated changes in subsurface elastic properties. With reconstruction guidance, forecasting and inversion can be achieved conditioned on historical frames and/or observational data. Meanwhile, due to the generative nature of the approach, we can quantify uncertainty in the prediction. Tests based on the Compass model show that the proposed method successfully captured the inherently complex physical phenomena associated with CO$_2$ monitoring, and it can predict and invert the subsurface elastic properties and CO$_2$ saturation with consistency in their evolution.

Figures (13)

• Figure 1: Workflow of the diffusion process and its reverse process, and the reconstruction guided generation for forecasting and inversion. The yellow box denotes the scaled learned score function. "History" denotes the given conditions $\mathbf{x}^{0:t}$ for the forecasting tasks to predict $\mathbf{x}^{t:T}$, where superscript stands time-lapse frames. "Observation" denotes partial physical variables of $\mathbf{x}^{0:T}$, e.g., RTM images, for the inversion of other variables.
• Figure 2: A multiphysics forward modeling framework starting with a permeability, which is converted from the baseline velocity model and then used to establish the CO$_2$ saturation given the constant injection using fluid flow simulation, which is in turn used to establish the changes in velocity and density based on rock physics modeling. Finally, the RTM images are obtained from the simulated data using the evolved velocity and density models.
• Figure 3: The training loss curve of the video diffusion model for multiphysics evolution.
• Figure 4: Uncodintional generation of multiphysics evaluation due to the CO$_2$ injection for two examples (a and b). Different frames denote different times with an interval of 3 years. Each row denotes a specific variable, including permeability, CO$_2$ saturation, velocity, density, and RTM images from top to bottom, respectively, while each column denotes the variables at different times.
• Figure 5: Forecasting generation (b) given three history frames (a), which are denoted by the red box. The solid color subfigures in (a) denote the unknown future states, which are the target of the forecasting generation. The rows represent properties described in Figure \ref{['fig:uncoditional']}.