A Deep Learning Framework to Forecast Spatial-Temporal Dynamics of CO₂ Mineralization in Reactive Rocks

Reactive rocks, such as basalt, are composed of a variety of minerals, with pyroxene, olivine, and plagioclase feldspar being among the predominant minerals. When exposed to carbon dioxide (CO₂)-charged waters, these rocks can undergo a series of reactions, leading to the formation of stable carbonates. These carbonates can store carbon for thousands of years, contributing to climate change mitigation. To better understand the interplay between CO₂ and brine in these reactive formations, numerical simulations are a very useful tool. However, simulating fluid flow in these reservoirs can pose significant computational challenges. The inherent reactivity of various minerals complicates the modeling process, leading to computationally expensive simulations. Therefore, the objective of this study is to develop a deep-learning workflow that can predict the changes in CO₂ mineralization over time and space in saline aquifers, offering a more efficient approach compared to traditional physics-based simulations. To achieve this, a numerical simulation model was created to replicate the CO₂ injection process in saline aquifers. The model was then sampled using the Latin-Hypercube method, considering various parameters related to petrophysics, geology, reservoir, and decision-making. These samples generated a comprehensive training dataset of approximately 700 simulation cases, forming the basis for training the UNet model, a type of convolutional neural network. The UNet models were trained, incorporating information on reservoir properties, well characteristics, and time, enabling the prediction of mineral precipitation at different spatial and temporal scales. During the training process, the root-mean-squared error (RMSE) was used as the loss function to prevent overfitting. Evaluation of the trained UNet model was performed using three error metrics: the normalized root mean square (NRMSE), mean absolute percentage error (MAPE), and the coefficient of determination (R²). The model achieved an R² value of 0.998 for training and 0.991 for testing, indicating its accuracy in predicting the evolution of mineral concentrations over time and space. The MAPE for all mappings was approximately 5%, demonstrating the effectiveness of the trained model. In terms of computational efficiency, the UNet model's prediction CPU time per case was remarkably fast, averaging only 0.2 seconds. This is significantly faster compared to the time required by the physics-based reservoir simulator, which took 21600 seconds per case. Thus, the proposed method not only provides accurate predictions comparable to physics-based models but also offers substantial computational time savings. The deep learning models developed in this study offer a computationally faster alternative to traditional numerical simulators for assessing mineralization trapping in geological carbon storage (GCS) projects, specifically concerning the mineral trapping mechanism.