Spotlight: Shell Accelerates CO2 Storage Modeling 100,000x Using NVIDIA Modulus

As the world faces the urgent need to combat climate change, carbon capture and storage (CCS) has emerged as a crucial technology for achieving net-zero emissions. The CCS technology—which involves capturing carbon dioxide (CO₂), either from industrial emissions or through direct air capture (DAC), and securely storing it in the subsurface—can drive much-needed decarbonization strategies and help achieve global climate targets.

The success of CCS technology depends on the careful selection of storage sites and injection schemes. Accurate predictions of CO₂ plume migration and pressure buildup over extended periods, often spanning hundreds of years, are essential for ensuring the safety and efficacy of storage sites. Finding the optimal setup requires assessing tens of thousands of configurations under varying subsurface conditions, well locations, and injection rates. However, conventional flow simulators, which are typically used for these predictions, are computationally intensive and thus limit the high-throughput screening of potential sites and injection schemes.

To address these challenges, Shell, in collaboration with NVIDIA, is leveraging cutting-edge technology through NVIDIA Modulus to enhance the efficiency and accuracy of CCS site screening processes.

Strategy to build a rapid screening tool for carbon capture and storage

The project employs machine learning (ML) models to enable rapid high-resolution modeling of subsurface CO₂ storage. This innovative approach not only accelerates the deployment of CCS technology but also enables more informed decisions about where and how to best store CO₂, ultimately contributing to the global effort to mitigate climate change. The expertise of Shell in the energy industry is combined with the leadership on NVIDIA in AI and the computational modeling space to build the technological advancements and breakthroughs to rapidly screen potential sites for CO₂ storage in a cost-effective and timely manner.

A flow chart showing various functionality centered around the CCS screening tool, including AI-based surrogate models, spatiotemporal prediction, improvised uncertainty assessment, and prior geology model operating conditions. — Figure 1. A CCS screening tool enables spatiotemporal modeling of CO₂ in subsurface reservoirs reflecting realistic geologies. AI-based surrogates provide unprecedented speedup enabling improved uncertainty assessments

A major challenge in identifying the most suitable sites for storage is assessing post-injection containment of CO₂ within a reservoir. This is because CO₂ migrates over time and potentially escapes from the reservoir, posing a risk to the environment. In addition, the reservoir pressure buildup caused by CO₂ injection needs to be carefully managed to avoid cracks in the geological layers, which are sealing the top of the reservoir and to prevent seismic hazards on the surface.

Researchers from Stanford University, California Institute of Technology, and Purdue University have shown that advanced AI-based surrogate models of subsurface CO₂ behavior substantially reduce computational costs relative to traditional numerical models, while preserving high levels of accuracy. To learn more, see U-FNO—An Enhanced Fourier Neural Operator-Based Deep-Learning Model for Multiphase Flow and Real-Time High-Resolution CO2 Geological Storage Prediction Using Nested Fourier Neural Operators.

Consequently, these models enable the examination of tens of thousands of injection configurations, facilitating more comprehensive and rapid screening of potential storage sites. To scale this research to industrial settings, Shell and NVIDIA jointly developed an AI framework based on Fourier neural operators (FNOs) to emulate the behavior of CO₂ in the reservoir. The software tool heavily leverages NVIDIA Modulus, an open-source framework for building, training, and fine-tuning Physics-ML models with a simple Python interface.

NVIDIA Modulus provides an extensive collection of neural network and neural operator architectures alongside convenience functions for setting up and scaling out training and inference pipelines. With the heavy lifting on the ML side being done, Modulus enables domain scientists and engineers to apply state-of-the-art ML techniques to their problems and scale them to massively parallel settings and deployment in production.

Results and applications

The model is trained on a comprehensive dataset generated from realistic subsurface parameters simulating 1,000 years post-injection. The fully trained model offers O(10⁵[AC1] ) computational speedup with minimal sacrifice in prediction accuracy. It was shown that just one site assessment justifies training such a model. The fully trained model can be applied to numerous screening tasks, amplifying the benefits of this approach.

To analyze the accuracy of the model, we focused on spatiotemporal distributions of mass of CO₂ ( $m_{CO_2}$ ), gas saturation ( $S_g$ ), and pressure buildup ( $\delta p$ ). As illustrated in Figure 2, the model’s prediction shows strong qualitative agreement with the ground truth for all three variables.

Figures depicting the simulation domain overlayed with pseudocolor plots of gas saturation, pressure buildup, and mass accumulation. — *Figure 2. Visual comparison of the fields for $S_g$ , $m_{CO_2}$ , and $\delta p$ [AC2] as predicted by the model with respective ground truth*

*Figure 2. Visual comparison of the fields for $S_g$ , $m_{CO_2}$ , and $\delta p$ [AC2] as predicted by the model with respective ground truth*

Figure 3a shows the mean absolute error (MAE) for the two solution variables of interest; MAE is a good metric for assessing the accuracy of the variables over the entire domain. In addition, $p_{90}$ [ $m_{CO_2}$ ], an indicator for the migration distance of the CO₂ plume from the injection location is monitored. Figure 3b provides the $R^2$ correlation plots of $p_{90}$ over time. For pressure buildup, global metrics are not informative. Instead, local, pointwise metrics for assessing the accuracy of the pressure predictions are explored. Histograms of maximum point-wise error in ( $\delta p$ ) across samples are shown in Figure 3c.

While monitoring CCS sites, the location where the maximum pressure occurs is of specific interest—typically near the injection well. Thus, the prediction from the model at the location where the true pressure is maximum (red circles in Figure 3d) is evaluated. To avoid location bias, we randomly select additional locations from the test set and assess predictions from the model at the same location (blue circles in Figure 3d). In both metrics, a $R^2$ score greater than 0.97 is observed, suggesting that the model would provide reliable predictions in most scenarios.

For more detailed information, see Fourier Neural Operator Based Surrogates for Storage in Realistic Geologies. The paper also presents super-resolution experiments and strategies for further improving the reliability of the predictions from the model, which is crucial while assessing actual geological sites.

Error metrics a) MAE for mCO2 and Sg, with MAE representing average in time b) R2 correlation plots of 90% plume mass migration distance p90. c) Maximum pointwise error in δp d) R2 correlation plots of δp at maximum pressure location corresponding to the test sample shown with red circles. Correlation plots δp for randomly selected locations within the active domain of the test sample shown with blue circles. — *Figure 3. Quantitative error metrics for gas saturation, pressure buildup, and plume mass migration*

Conclusion

Shell, in collaboration with NVIDIA, has developed a machine learning model based on the Fourier neural operators for real-time, high-resolution simulation of CO₂ plume migration. This model is trained on an extensive dataset derived from realistic subsurface parameters. During inference, we observe a speedup of O(10⁵[AC3] ) over traditional numerical simulators of CO₂ flow fields with minimal reduction in accuracy. Along with fast surrogate models, we present and assess several physics-based accuracy metrics that are relevant for assessing and monitoring CCS sites.

Additionally, we propose several strategies—outlier detection, enforcing mass conservation—to enhance the reliability of the model’s predictions, which is vital when evaluating actual geological sites. Our work scales up scientific machine learning models to realistic systems that are more consistent with real-life subsurface reservoirs and builds the foundation for a first-of-its kind advanced screening tool for subsurface CCS applications.