Physics-Informed Graph Neural Networks for Multi-Resolution CO₂ Saturation Estimation in Subsurface

Objectives:

In this study, we aim to develop a data-driven petrophysical inversion technique in the context of CO₂ sequestration. By integrating reservoir flow simulation, petroelastic modeling, and Graph Neural Networks (GNNs), CO₂ saturation can be estimated within models with multi-grid resolutions. The goal is to enhance accuracy and resolution adaptability in predicting CO₂ plume behavior in subsurface geological formations, thus improving carbon capture and storage (CCS) strategies.

Methodology:

We generated 100 2-dimensional synthetic reservoir models using a sequential indicator simulation algorithm for facies simulation, each populated by heterogeneous porosity and permeability. Flow simulations were conducted for 11 years using a central well with a constant injection rate. Petroelastic modeling was then performed to compute changes in P-wave and S-wave velocities and density every six months. The models were resampled to mimic a varying resolution scenario, with higher resolution near the well. A GNN model handled multi-resolution inputs and outputs, representing each grid as a node linked to its nearest eight neighbors, using direction and distances as edge attributes.

Results, Observations, and Conclusions:

The integrated modeling approach successfully predicted CO₂ plume migration within geological formations, demonstrating high predictive accuracy and robustness. Petroelastic modeling revealed significant changes in reservoir properties such as P-wave and S-wave velocities and density due to CO₂ injection. The Graph Neural Network (GNN) model, optimized through hyperparameter tuning, effectively utilized these changes to predict CO₂ saturation with a Mean Squared Error (MSE) of 0.0217 and a Coefficient of Determination (R²) of 0.981, confirming its high reliability in practical scenarios. In comparison, the Multilayer Perceptron model (MLP) achieved an MSE of 0.0260 and an R2 of 0.9695, processing data without considering spatial connections, underscoring the GNN's superior computational efficiency and spatial data integration. Furthermore, visual assessments confirmed the model’s accuracy, closely aligning predicted and actual CO₂ saturation levels, especially in dynamically changing reservoir zones. The study concludes that combining static property modeling, flow simulation, petroelastic modeling, and GNNs provides a valuable tool for enhancing CO₂ sequestration strategies, improving the prediction accuracy of CO₂ behavior in the subsurface, and significantly advancing CCS technologies.

Novel/Additive Information:

Our work leverages Graph Neural Networks (GNNs) to predict changes in CO₂ saturation from elastic properties, integrating flow dynamics with petroelastic modeling and deep learning via adaptive meshing grids. This novel approach addresses the limitations of conventional neural networks in adapting to mesh variations. Our project uniquely targets the complex challenges of CO₂ monitoring, advancing sequestration monitoring technologies by bridging seismic monitoring and dynamic flow simulation, providing a tool to predict CO₂ saturation from elastic properties.