Carbon capture and storage (CCS) has become a key element in achieving climate neutrality goals. However, one of the major technical challenges is predicting how the land surface will respond to CO₂ injection into the subsurface. Surface deformations can compromise reservoir integrity and generate operational risks. Traditionally, these predictions require extremely costly numerical simulations, with large-scale geomechanical models that demand hours of computing and do not always generalize well to new scenarios. This is where CarbonNet emerges: a novel approach based on computer vision that analyzes subsurface geometry images to estimate surface displacement. Instead of solving complex equations, CarbonNet trains deep learning models directly on visual representations of underground layers, achieving fast and accurate predictions. This article explores the technology behind CarbonNet, compares different architectures (CNN, ResNet, ResNetUNet, LSTM, and Transformer), and discusses how this solution can be integrated into real CCS projects, with support from technology services such as those offered by Q2BSTUDIO.
The core problem is predictive in nature: given a set of images representing subsurface geometry (porosity, permeability, geological structures), we need to forecast the displacement field on the surface. This is essential for monitoring storage integrity and preventing leaks. Traditional solutions rely on finite element or finite difference methods, which are accurate but computationally prohibitive when exploring multiple injection scenarios. CarbonNet proposes a radical alternative: turning the problem into supervised machine learning, where input images are subsurface properties and the output is a surface displacement image. To this end, a synthetic dataset was generated through geomechanical simulations, creating input-output pairs that allowed training convolutional and recurrent networks.
In the static mechanics scenario, where deformation does not vary over time, three architectures were evaluated: classic CNN, ResNet, and ResNetUNet. Results showed that ResNetUNet outperforms the others due to its ability to preserve fine spatial details through skip connections, combining high- and low-level features. This is crucial because surface deformations often exhibit localized patterns that a fully convolutional network might miss. For the transient case, where injection occurs over time and deformation evolves, LSTM and Transformer were employed. Surprisingly, LSTM offered comparable performance to Transformer, a relevant finding because it suggests that for relatively short geomechanical sequences, LSTM's long-term memory is sufficient, without the need for the attention complexity of the Transformer. CarbonNet thus demonstrates that computer vision can dramatically reduce computational cost without sacrificing accuracy.
The implementation of CarbonNet is not just an academic advancement; it has direct business implications. Energy and technology companies operating in CCS need agile tools for decision-making. This is where a technology partner like Q2BSTUDIO can make a difference. With expertise in custom software development, artificial intelligence, cybersecurity, cloud computing (AWS/Azure), and business intelligence (Power BI), Q2BSTUDIO offers the capabilities needed to industrialize CarbonNet. For example, a CCS project might require a cloud platform that deploys computer vision models at scale, with secure authentication and BI dashboards to visualize predictions in real time. Furthermore, integrating AI agents (intelligent assistants) would allow engineers to query hypothetical scenarios using natural language, accelerating predictive analysis.
CarbonNet represents a paradigm shift in computational geomechanics. By harnessing the power of neural networks and computer vision, it eliminates the need to solve differential equations every time a new injection point is evaluated. Trained models can generalize to unseen geological configurations, provided the synthetic dataset is sufficiently representative. This opens the door to rapid sensitivity studies, injection parameter optimization, and continuous monitoring. From a business perspective, adopting these techniques reduces simulation times from weeks to minutes, translating into significant operational cost savings and better responsiveness to plan deviations.
Of course, challenges remain. Prediction quality heavily depends on the fidelity of the synthetic dataset, and generalization to real fields requires validation with field data. Moreover, model interpretability is still an active research area: geomechanical engineers need to understand why the model predicts a certain deformation, not just the numerical value. Here, combining explainable AI (XAI) techniques with BI dashboards can provide transparency. Q2BSTUDIO, with its Business Intelligence and process automation offerings, can help build these interpretation layers, integrating CarbonNet predictions into decision support systems.
In conclusion, CarbonNet demonstrates that computer vision is a viable and powerful tool for predicting surface deformations in CCS projects. Its ability to process subsurface images and generate displacement maps in seconds makes it a key enabler for real-time monitoring and agile decision-making. Companies seeking to implement similar solutions can benefit from Q2BSTUDIO's service ecosystem, ranging from custom application development to cloud security and artificial intelligence. CarbonNet is not just a concept; it is the beginning of a new generation of data-driven geomechanical tools, where speed and accuracy go hand in hand.
