CO₂ capture and storage is a viable solution in the effort to mitigate global climate change. Deep saline aquifers, in particular, have emerged as promising storage options, owing to their vast capacity and widespread distribution. However, the task of proficiently monitoring and simulating CO₂ behavior within these formations poses significant challenges. To address this, we introduce the physics-constraint neural network for CO₂ storage (CO₂PCNet), a model specifically designed for simulating and monitoring CO₂ storage in deep saline aquifers during injection and post-injection periods. Recognizing the significant challenges in accurately modeling the distribution and movement of CO₂ under varying permeability conditions, the CO₂PCNet integrates the principles of physics with the robustness of deep learning, serving as a powerful surrogate model. The architecture of CO₂PCNet starts with an encoder that adeptly processes spatial features from overall mole fraction (z_CO2) and pressure fields (P_l), capturing the complex dynamics of a CO₂ trajectory. By incorporating permeability information through a conditioning step, the network ensures a faithful representation of the influences on CO₂ behavior in subsurface conditions. A ConvLSTM module subsequently discerns temporal evolutions, reflecting the real-world progression of CO₂ plumes within the reservoir. Lastly, the decoder precisely reconstructs the predictive spatial profile of CO₂ distribution. CO₂PCNet, with its integration of convolutional layers, recurrent mechanisms, and physics-informed constraints, offers a refined approach to CO₂ storage simulation. This model offers the potential of utilizing advanced computational methods in advancing CCS practices.

CO₂ sequestration and storage in deep saline aquifers is a promising technology for mitigating the excessive concentration of the greenhouse gas in the atmosphere. However, accurately predicting the migration of CO₂ plumes requires complex multi-physics-based numerical simulation approaches, which are prohibitively expensive due to highly nonlinear coupled governing equations and uncertainties in heterogeneous spatial parameter distributions. To address this challenge, we developed an end-to-end deep learning workflow employing encoder–decoder architectures with residual network (ResNet) to efficiently predicts the spatial–temporal evolution of the solution CO₂-brine ratio (R_s) and gas saturation (S_g) – the two essential tasks for quantifying the amount of trapped CO₂ – given heterogeneous permeability fields as input. Specifically, we introduce a general multi-task learning with consistency (MTLC) framework to simultaneously predict R_s and S_g. The MTLC model leverages related tasks with less computational expensive labeled datasets to improve generalization ability. In our study, predictions for multiple tasks from the same permeability realization are not independent but expected to be consistent, as the proposed framework utilizes data-driven cross-task consistency constraints to augment learning of related tasks. Our deep learning model is trained based on physical trapping mechanisms, which play a dominant role in the CO₂ migration process. The results demonstrate that the MTLC model with joint learning yields more accurate predictions and improved generalization for predicting CO₂ migration in several test cases. Furthermore, our workflow is 10⁵ times faster than a high-fidelity physics-based numerical simulator, making it a viable alternative for field-scale applications.