Harnessing the power of machine learning for the optimization of CO2 sequestration in saline aquifers: Applied on the tensleep formation at teapot dome in Wyoming

The sequestration of carbon dioxide in deep saline aquifers offers a highly promising approach to mitigate CO₂ emissions resulting from the fossil fuels industry. The practicality of this solution is mostly dependent upon the CO₂ trapping efficiency, which governs the capacity of the aquifer for CO₂ storage. Generally, the compositional reservoir simulation is employed to evaluate the trapping efficiency, but is computationally expensive, particularly when adjusting parameters necessitates hundreds of simulations. In this paper, A machine learning (ML) proxy model was used to address these difficulties for fast evaluation and optimization of the residual and dissolution trapping mechanisms in the Tensleep sandstone formation in the Teapot Dome Field, located in Wyoming, USA. The initial CO₂ storage capacity of the aquifer was calculated to be 17.7 thousand tons. In the simulation model, several injection wells were placed in the high porosity regions and CO₂ injection was simulated over duration of 10 years, followed by a 90-year post-injection period. The injection rates were optimized to maximize the overall trapping efficiency, which takes into account both residual and solubility indices. In order to construct a dataset for machine learning-based proxy models, the Latin hypercube sampling technique was adopted to generate 100 simulation runs that varied operating constraints: maximum injection rate, maximum injection pressure, and number of injection wells. The Radial Basis Function-Artificial Neural Network (RBF-ANN) was specifically trained to accurately determine the most appropriate injection rates by identifying complex and non-linear correlations within the data. The total CO₂ trapping effectiveness by the RBF-ANN model was enhanced from 75% to 83%, accompanied by an insignificant increase in the leakage index from 0.64% to 1.3%. The results indicated that machine learning proxy modeling offers a rapid and accurate approach to optimize the storage of CO₂ in saline aquifers. Through the reduction of CO₂ emissions, this method significantly improves the viability of large-scale sequestration projects, so making a valuable contribution to climate change mitigation.