Gradient-guided Convolutional AutoEncoder for predicting CO₂ storage in saline aquifers with multiple geological scenarios and well placements.

The article focuses on the development and evaluation of a Gradient-guided Convolutional AutoEncoder (GCAE) for predicting CO₂ sequestration in deep saline aquifers, specifically targeting the mole fraction of CO₂ dissolved in the aqueous phase. By incorporating spatial gradient differential operators as physical prior knowledge into the neural network's loss function, GCAE enhances prediction accuracy, data stability, and generalization capability compared to a purely data-driven Convolutional AutoEncoder (CAE). The study demonstrates GCAE's superior performance across homogeneous and heterogeneous geological models, including scenarios with varied well placements, highlighting its improved physical interpretability and robustness. While GCAE shows promise as a surrogate model for CO₂ storage processes, the authors note it complements rather than replaces classical numerical methods for solving partial differential equations.