Machine Learning for Prediction of CO2 Foam Flooding Performance

Abstract In a move towards development of sustainable and efficient hydrocarbon production, the industry looks forward to the deployment of carbon neutral and even carbon negative solutions. Accordingly, CO2 EOR is a viable option to improve recovery and has been applied in mature fields for over four decades. The downsides of poor sweep efficiency linked to viscous fingering and gravity segregation can be sorted through generation of CO2 foams in the reservoir. This work proposes the utilization of machine learning techniques, to predict foam flood performance which will thereby aid in optimization of laboratory core-flood experiments. This work is based upon consumption of large set of existing laboratory data collected from literature, amounting to more than 200 data points. The dataset reports core oil recovery factor as a function of three reservoir parameters including porosity, permeability, initial oil saturation. While injected foam volume and total pore volume are also considered. Furthermore, the data records contain experiments for various foaming agent types which are catered for during the machine learning model development through the implementation of numerical tags. The input data is then divided in training subset for development of XGBoost model, complemented by integration of exhaustive grid search and k-fold cross validation techniques. Subsequently, the testing subset is reserved to measure efficacy of the developed model. The model development process involves tuning of machine learning algorithm hyperparameters which control the resultant accuracy, while at the same time it is ensured that the issue of model overfitting is avoided. Testing of the established model is carried out through an array of statistical measures including the R2 and RMSE values. The proposed model is compared with actual experimental data. The machine learning model can achieve high accuracy in predictive mode for the output parameters. Through statistical error analysis performance measurement, it is observed that the machine learning model can predict CO2 foam flood performance with high R2 of around 0.99 and low errors. The excellent accuracy of the XGBoost model is credited to the complex processing involved with intelligent algorithms that can discover underlying relationships among the input variables.