Abstract

Mitigating rising atmospheric CO₂ levels is critical to addressing climate change, and geological carbon storage in saline aquifers is a promising solution. This study develops and evaluates three deep learning modelsdeep neural network (DNN), gated recurrent unit (GRU), and recurrent neural network (RNN)to predict the residual trapping index (RTI) and solubility trapping index (STI) via diverse reservoir datasets. The GRU model outperformed both DNN and RNN, particularly in handling long-term dependencies and minimizing error, while DNN showed robust accuracy through effective modeling of complex nonlinear relationships. In contrast, RNN exhibited challenges with gradient instability, affecting its prediction performance. Sensitivity analysis highlighted the significance of input variables like post injection and injection rate, with DNN showing greater dependence on these features. Excluding these variables reduced predictive accuracy, especially for RTI. These findings suggest that GRU is the most effective for predicting CO₂ trapping, offering a valuable tool for optimizing carbon storage strategies and supporting global carbon reduction efforts.