CarbonNet: How Computer Vision Plays a Role in Climate Change? Application: Learning Geomechanics from Subsurface Geometry of CCS to Mitigate Global Warming

Technical Summary: CarbonNet – Computer Vision for Geomechanics in CCS

Problem Statement
Carbon Capture and Sequestration (CCS) is a critical technology for mitigating global warming, involving the injection of supercritical CO₂ into deep brine aquifers or abandoned oil reservoirs. A primary risk associated with this process is surface upheaval (land displacement) caused by excessive fluid pressure, which can damage infrastructure and affect human safety. Predicting the spatial distribution and temporal evolution of this surface deformation is essential for hazard avoidance. However, traditional numerical simulations (e.g., Finite Element Method) are computationally prohibitive due to the large spatial and temporal scales of CCS projects. Furthermore, existing deep learning approaches often struggle with material heterogeneity, coupling between fluid flow and mechanics, or are limited to specific geological settings. This project addresses the challenge of efficiently predicting land surface displacement under varying subsurface geometries using computer vision techniques.

Methodology
The authors formulate the problem as an image-to-image prediction task for static mechanics and a video prediction task for transient mechanics. The input consists of 2D images representing subsurface geometry (specifically, a shale layer with varying dip angles embedded between two permeable layers), and the output is the corresponding surface displacement field.

1. Dataset Generation:

   • Static Model: 10,000 samples were generated using FEniCS (FEM-based simulation). Inputs are geometry images with dip angles ranging from 0° to 45°. Labels are vertical displacement ($u_y$) values at 1,250 points on a 50×25 grid.
   • Transient Model: 3,000 samples were generated using coupled hydro-mechanical simulations. Inputs are geometry images, and labels are time-series data of vertical displacement at 40 points on the ground surface over 1,000 time steps. Data sequences were created using sliding windows (5 steps to predict the next) and normalized via min-max scaling.

2. Static Mechanics Models (Image Prediction):

   • Baselines: Standard CNN and ResNet (ResNet18) architectures were tested. These models downsample images to extract features but were found to lose critical spatial information, leading to inaccuracies in heterogeneous regions.
   • Proposed Model (ResNetUNet): A custom architecture combining ResNet and UNet. The encoder utilizes ResNet blocks to capture complex physical laws and heterogeneous material distributions, while the decoder employs UNet-style upsampling with skip connections to recover spatial details and prevent information loss.

3. Transient Mechanics Models (Video/Time-Series Prediction):

   • Feature Extraction: Both models utilize the pre-trained ResNetUNet to extract features from the input geometry image.
   • LSTM: A Long Short-Term Memory network (stacked 4 LSTM cells) processes the sequence of displacement values (size 40) to predict the next time step.
   • Transformer: A Transformer encoder-decoder architecture processes the sequence. It uses self-attention mechanisms to capture long-term dependencies, theoretically overcoming the sequential information loss inherent in LSTMs. The encoder and decoder each consist of four layers with a model dimension of 16 and 8 attention heads.

Key Results

• Static Mechanics: The ResNetUNet model significantly outperformed both CNN and ResNet baselines. Visual comparisons showed ResNetUNet predictions were nearly identical to the physics-based ground truth, whereas CNN and ResNet exhibited information loss and lower accuracy, particularly in heterogeneous regions. Quantitatively, ResNetUNet achieved a validation MSE of $7.33 \times 10^{-8}$ and MAE of $2.00 \times 10^{-4}$, orders of magnitude lower than the baselines.
• Transient Mechanics: Both LSTM and Transformer models successfully predicted the temporal evolution of surface displacement. The Transformer model demonstrated superior performance with a lower validation MSE ($3.86 \times 10^{-9}$) and MAE ($1.50 \times 10^{-6}$) compared to the LSTM. While the LSTM showed promising potential, the Transformer's self-attention mechanism allowed for better handling of the high-dimensional sequence data (40 points), fitting the ground truth nearly perfectly.
• Training Dynamics: All models were trained without cross-validation due to the abundance of synthetic data. Loss histories indicated no signs of overfitting or underfitting.

Significance and Claims
The paper claims to present a state-of-the-art approach that leverages computer vision intelligence to solve complex coupled geomechanics and fluid flow problems. The primary significance lies in generalizing the solution to arbitrary input geometry images, moving beyond models restricted to specific geological formations.

• Efficiency: The approach offers a computationally efficient alternative to traditional numerical simulations, enabling faster decision-making for CCS project planning.
• Architecture Efficacy: The study demonstrates that the ResNetUNet architecture is particularly effective for heterogeneous static mechanics problems by preserving spatial information through skip connections.
• Temporal Modeling: The work validates the applicability of Transformer models in geomechanical time-series prediction, showing they can outperform traditional recurrent networks (LSTM) in capturing long-term dependencies in physical systems.
• Future Impact: The authors posit that these models can inform decision-making in carbon storage surveillance to mitigate climate change. Future work includes testing robustness with generative adversarial networks (GANs) for more complex geometries and deploying an alpha version of a web application for public access.