A recurrent neural network surrogate model with few-shot learning strategy for CO2 storage in deep subsurface saline aquifer

Carbon dioxide (CO₂) storage in deep subsurface saline aquifer is a crucial technology for achieving long-term and safe CO₂ storage. Conducting numerical simulations of deep subsurface saline aquifer CO₂ storage projects to evaluate engineering plans before implementation demands significant computational resources. The application of deep learning surrogate models can effectively enhance simulation efficiency and reduce the consumption of computational resources. However, constructing a deep learning surrogate model in practical tasks typically requires a substantial number of simulations to gather sufficient training samples. Therefore, the advantages of reduced resource consumption offered by deep learning surrogate model are limited compared to the deployment cost associated with the simulation process. To address these deployment costs, this study proposes a deep learning surrogate model, 3D ConvLSTM ResU-Net++ (hereafter referred to as 3D CRU-Net++), optimized using the Reptile algorithm. By employing a few-shot strategy based on Reptile, this model can simulate the CO₂ storage process under varying reservoir conditions with relatively low computational resource requirements. The surrogate model was trained using diverse geological storage scenarios, and a target task was selected to validate its effectiveness. The experimental results show that the model optimized using Reptile demonstrated a 70% reduction in RMSE, a 60% reduction in MAE, and a 7% improvement in R². These results suggests that Reptile-optimized model achieves higher accuracy with the same number of samples compared to the standard model. This study shows that the Reptile algorithm significantly reduces the number of samples required for training surrogate models, thus facilitating the application of these models in various CO₂ sequestration simulations, such as varying reservoir conditions, injection well optimization, and uncertainty analysis.