Study on utilization of machine learning techniques for geological CO2 sequestration simulations

Abstract

In Geological Carbon Sequestration (GCS), CO2 is injected and stored in geological formations. Conducting these studies through experiments could be highly impractical both economically and scientifically. Employing conventional numerical techniques was proven to be computationally intensive and slow with complex geometrical models. In this work, a study is conducted to integrate traditional numerical simulation results with machine learning techniques to forecast the futuristic trend of output parameters. Primary simulations are performed using multiphase flow modelling for the entire geological time scale. Then the time series neural network is used by considering CO2 sequestration parameter values as input and target data to forecast the trend of output parameters during the post-injection time. Both recurrent neural network models have shown a reasonable forecast of the output variables. Among the different training algorithms used, the Levenberg-Marquardt (LM) algorithm has given a good prediction for the output variable; the results are validated with the RMSE and R-squared values for obtained values. The R2 and RMSE values of the NAR model for structural trapping are 0.9801 and 0.0515, respectively, and for residual trapping, they are 0.9805 and 0.0506, respectively. This work will provide the initial understanding of integrated machine learning techniques in the GCS analysis for a heterogeneous reservoir model.