Computer vision model predicts land shifts from carbon storage, cutting simulation costs

Carbon capture and sequestration (CCS) is widely considered essential for achieving carbon-neutral societies, yet the technology faces significant practical hurdles. Computing the geomechanical effects of injecting CO₂ underground requires expensive simulations, and existing models often struggle to adapt to new geological conditions. A new approach tackles these limitations by training computer vision systems to learn the relationship between subsurface rock formations and surface displacement.

The Challenge with Current Methods#

Traditional methods for predicting how the ground responds to carbon injection involve running detailed physics-based simulations that demand substantial computational resources. Additionally, pre-trained models trained on one site's geology frequently fail to generalize to different subsurface conditions, limiting their practical reusability.

Training Models from Images#

Rather than relying on expensive numerical solvers, researchers developed machine learning models that work directly from subsurface geometry images—essentially learning the geomechanical response visually. For scenarios where conditions don't change over time (static mechanics), they tested convolutional neural networks (CNN), ResNet, and ResNetUNet architectures. For cases where displacement evolves as CO₂ spreads (transient mechanics), they deployed LSTM and transformer models adapted for video prediction tasks.

Results and Performance#

In static mechanics problems, ResNetUNet outperformed the other architectures due to its encoder-decoder design. For time-dependent scenarios, LSTM and transformer models achieved comparable accuracy, suggesting either approach could be viable depending on deployment constraints. These findings indicate that direct image-based learning can substitute for costly physics simulations while remaining accurate enough to guide CCS project decisions.

Implications for Carbon Management#

If validated at scale, this method could accelerate permitting and site selection for CCS facilities by providing faster, cheaper displacement forecasts. The work demonstrates how computer vision—traditionally associated with autonomous systems and media analysis—can be repurposed for environmental engineering and climate mitigation.