Rapid Forecasting of Trapped and Movable Carbon Dioxide in Saline Aquifers using LSTM-based Sequence-to-Sequence Learning

Abstract Geological carbon storage (GCS) will play a crucial role in reducing greenhouse gas emissions. Deep saline aquifers are considered suitable sites for geological carbon storage due to their accessibility, storage capacity, and containment efficiency. The paper develops a deep-learning-based rapid forecasting workflow to visualize the temporal evolution of the trapped and movable CO2 in the subsurface aquifer during a geological carbon storage operation. Rapid forecasting enables agile decision-making by providing timely insights into rapidly changing environments. This study presents the application of an LSTM-based Seq2Seq model for predicting the temporal variations in proportions of residually trapped, solubility trapped (dissolved), and movable CO2. A dataset comprising 1600 simulations of CO2 evolution in saline aquifer under various geological and engineering parameters was utilized as training and testing dataset. The LSTM-based Seq2Seq model was trained and tested to forecast the dynamic temporal variations in residually trapped and solubility trapped CO2 mass and the movable CO2 mass fractions over a period of 80 years, involving injection phase followed by monitoring phase. The prediction outcomes demonstrate that the LSTM-based Seq2Seq model not only attains a Coefficient of Determination (R2) of up to 0.99 but also requires merely 0.35 milliseconds to forecast the movable and trapped CO2 mass, which is six orders of magnitude computational speed-up as compared to conventional simulator.