Search for a command to run...
• Multi-resolution deep-learning model allows training using upscaled reservoir model. • 90 % reduction in training data generation cost using upscaled reservoir model. • Proposed workflow accelerates optimization by orders of magnitude. • CCS well control at Illinois Basin Decatur Project has been optimized. • Both surface injection and zonal rate allocation optimization are conducted. The injection of CO₂ into subsurface formations entails various risks, necessitating a comprehensive evaluation of several factors such as seismic activity, seal integrity, and CO₂ leakage. Therefore, a CO₂ sequestration project involves multiple objectives which potentially exhibit trade-offs. Traditional multi-objective optimization frameworks require hundreds of forward simulations, which are computationally prohibitive for large-scale field application. In this study, we propose a deep learning-based workflow to efficiently optimize well control in CO₂ sequestration projects including surface injection and zonal rate allocation, enabling scalability to large-scale field applications. We developed a multi-resolution deep learning model based on the Fourier Neural Operator (FNO), which provides super-resolution capability. This capability allows the model to be trained on coarse-scale simulations while accurately predicting fine-scale reservoir pressure and saturation responses from permeability and injection schedules. As a result, data generation costs are substantially reduced, significantly lowering the overall cost of developing deep learning models. The original FNO architecture was modified to improve predictive accuracy across spatial resolutions, resulting in the proposed multi-resolution FNO model. This model functions as a data-driven proxy integrated with a multi-objective genetic algorithm to optimize CO₂ injection control, effectively balancing pressure management and storage efficiency. The power and efficiency of our approach are demonstrated on both synthetic and field applications, including a large-scale CO₂ injection at the Illinois Basin Decatur Project. Application to the synthetic model demonstrates the superior predictive performance of the developed multi-resolution FNO across coarse to fine-scale properties. For the field application, coarse-scale training data reduces training data generation cost by 90%, while the FNO-based proxy accurately predicts fine-scale pressure and saturation distribution, which are verified against a commercial reservoir simulator. The multi-objective optimization workflow, implemented using the FNO-based proxy model, achieves substantial improvements across multiple objectives while delivering performance orders of magnitude faster than traditional simulation-based approaches. We applied this workflow to CO₂ sequestration scenarios, including balancing pressure buildup with CO₂ injection amount, as well as optimizing surface and zonal injection rate allocations. This work introduces a novel multi-resolution FNO-based proxy model, applied to CO₂ injector control optimization. By combining FNO’s super-resolution capability with coarse-scale models, training data generation costs are greatly reduced. The proxy model accelerates forward simulations by orders of magnitude and enables efficient evaluation of multiple optimization scenarios for large-scale field applications.
Published in: International journal of greenhouse gas control
Volume 152, pp. 104633-104633