Uncertainty Quantification of Fluid Leakage and Fault Instability in Geologic CO2 Storage

Hongfang Lü, Lluis Salo-Salgado, Youssef Marzouk, Rubén Juanes

TL;DR

This paper introduces a deep-learning-based framework to predict the risks of fault reactivation and fluid leakage in geologic $CO_2$ storage. By combining a stochastic geological modeling tool (PREDICT) with a dynamical surrogate model called Flow Map Learning (FML), the researchers achieved a ten-fold speedup in uncertainty quantification while accurately capturing complex, multi-modal risk distributions.

What Problem Does This Paper Address?

Geologic $CO_2$ storage (GCS) is a critical tool for hitting net-zero targets, but it comes with "underground" risks. Injecting $CO_2$ increases reservoir pressure, which can cause faults to slip (triggering micro-seismicity) or allow brine and $CO_2$ to leak into freshwater aquifers.

The challenge lies in uncertainty. We rarely know the exact internal structure or permeability of a fault thousands of meters underground. To account for this, engineers usually run "Monte Carlo" simulations—running the same model thousands of times with different parameters to see the range of possible outcomes. However, high-fidelity geomechanical simulations are computationally "prohibitive," taking hours or days for a single run. This makes thorough risk assessment nearly impossible in real-time or on a budget.

Methods

The authors propose a three-stage pipeline to bypass the computational bottleneck:

  1. Stochastic Fault Modeling (PREDICT): Instead of assuming a uniform fault, they use the PREDICT methodology to generate realistic fault permeability tensors. This accounts for "clay smears"—discontinuous layers of clay that create complex, anisotropic flow paths (where fluid moves easier in one direction than another).
  2. Flow Map Learning (FML): This is the core innovation. Rather than training a Deep Neural Network (DNN) to predict the entire 3D pressure and stress field (which is data-heavy and slow), they train the model to learn the evolution of specific Quantities of Interest (QoIs):
  3. Maximum Slip Tendency ($T_{s,max}$): How close the fault is to sliding.
  4. Maximum Pressure Buildup ($\Delta p_{max}$): The stress on the caprock.
  5. Total Brine Leakage ($\Delta M_b$): The volume of displaced salt water escaping the reservoir.
  6. Dynamical Surrogates: The FML model acts as a time-marching numerical integrator. It uses "memory" (past states) to predict the next state, allowing it to forecast risks over varying time horizons without re-running the physics simulator.

Datasets

The study utilized synthetic datasets generated via the MATLAB Reservoir Simulation Toolbox (MRST):

  • Base Population: 1,000 high-fidelity simulations of a 2D plane-strain model featuring a storage aquifer, caprocks, and a normal fault.
  • Training Set: 100 samples selected from the population using a modified space-filling design. This ensures the training data covers the edges of the parameter space (extreme cases) rather than just the average cases.
  • Parameters: Uncertain variables included fault permeability ($k_f$), reservoir permeability ($k_s$), initial stress ratio ($\lambda$), porosity ($\phi_s$), and rock compressibility ($c_r$).

Key Insights and Ideas

  • QoIs over Full Fields: The "Aha!" moment is that for risk management, you don't need to know the pressure at every single coordinate. By focusing the AI on low-dimensional QoIs, the model becomes much cheaper to train and more accurate with less data.
  • The Power of Memory: By incorporating "memory steps" ($n_M$) into the DNN, the surrogate model can capture complex physical dependencies that a simple "input-to-output" mapping would miss.
  • Space-Filling vs. Random Sampling: The authors found that how you pick your 100 training simulations matters immensely. Random sampling often misses "bifurcations" in the data (where a small change in parameters leads to a totally different outcome), while space-filling designs capture these critical transitions.
  • Anisotropy Matters: The study highlights that ignoring the directional nature of fault permeability leads to a significant underestimation of leakage and instability risks.

Key Results

  • Accuracy: The FML surrogate successfully reproduced the 10th, 50th, and 90th percentiles of the risk distributions, matching the "ground truth" from expensive physics-based simulations.
  • Speed: The computational cost of building and running the surrogate (including generating the 100 training samples) was one order of magnitude lower than running the full 1,000-sample Monte Carlo suite.
  • Sensitivity Analysis: Using Sobol' indices, the authors showed that:
  • Fault Stability is dominated by the initial stress ratio ($\lambda$).
  • Pressure Buildup is most sensitive to fault permeability ($k_f$).
  • Brine Leakage is primarily driven by reservoir porosity ($\phi_s$).

Limitations and Future Work

  • 2D Constraints: The current model is 2D. Real-world faults are 3D structures with even more complex connectivity.
  • Fixed Injection Scenarios: The study focused on a constant injection rate. Future work could involve using the surrogate for optimal control, where injection rates are adjusted in real-time to maximize storage while keeping slip tendency below a safety threshold.
  • Data Requirements: While 100 samples is "cheap," it still requires significant initial compute. Exploring "active learning" (where the model decides which simulation to run next to learn the most) could further reduce this cost.

Summary updated

Abstract

Geologic CO$_2$ storage is an important strategy for reducing greenhouse gas emissions to the atmosphere and mitigating climate change. In this process, coupling between mechanical deformation and fluid flow in fault zones is a key determinant of fault instability, induced seismicity, and CO$_2$ leakage. Using a recently developed methodology, PREDICT, we obtain probability distributions of the permeability tensor in faults from the stochastic placement of clay smears that accounts for geologic uncertainty. We build a comprehensive set of fault permeability scenarios from PREDICT and investigate the effects of uncertainties from the fault zone internal structure and composition on forecasts of CO$_2$ permanence and fault stability. To tackle the prohibitively expensive computational cost of the large number of simulations required to quantify uncertainty, we develop a deep-learning-based surrogate model capable of predicting flow migration, pressure buildup, and geomechanical responses in CO$_2$ storage operations. We also compare our probabilistic estimation of CO$_2$ leakage and fault instability with previous studies based on deterministic estimates of fault permeability. The results highlight the importance of including uncertainty and anisotropy in modeling of complex fault structures and improved management of geologic CO$_2$ storage projects.
BibTeX
@inproceedings{lu2025uncertainty,
  author = {Hongfang Lü and Lluis Salo-Salgado and Youssef Marzouk and Rubén Juanes},
  title = {{Uncertainty Quantification of Fluid Leakage and Fault Instability in Geologic CO2 Storage}},
  booktitle = {Water},
  year = {2025},
  doi = {10.1029/2024wr039275},
}
Figure