Optimization of CO2 Foam Flooding Performance Enhanced by Nanoparticles Using Machine Learning Techniques

Abstract

This thesis explores how nanoparticle-enhanced CO₂ foam flooding can be optimized by using machine learning, with half-life of the foam chosen as the most significant parameter of foam stability. Even though the CO₂ foam flooding is a promising technique of enhanced oil recovery because of its enhanced mobility control and sweep efficiency, its operation is often hindered by the instability of the foam at varying formulation and working conditions. Due to nonlinear behaviour and many interacting variables which control foam stability, conventional linear techniques are not adequate to predict and optimize accurately. To overcome this, a predictive model was established based on experimental data on 31 published studies on nanoparticlebased CO₂ foam systems. The original 543 observations data set was refined and narrowed to 355 samples with the input variables including salinity, temperature, injection rate, type of nanoparticle, size, and concentration, including the type and concentration of the surfactant. Exploratory analysis showed that there were weak linear correlations, which implies that there were nonlinear interactions. Various machine learning models have been designed and evaluated based on the R², RMSE, MAE and the 5-fold cross-validation algorithm, where the results indicate that nonlinear models outperform the others. The CatBoost model had the highest predictive accuracy (R² = 0.826, RMSE = 6.837, MAE = 4.111) and was further analysed. Optimization, sensitivity analysis, and interpretability models such as partial dependence plots, ICE plots, and SHAP analysis, demonstrated that foam stability is regulated by optimal parameter ranges as opposed to monotonic trends. It was found that the most significant variable was the size of nanoparticles, then the concentration of the nanoparticle, and the rate of injection, followed by a weaker and more context-dependent effect of temperature. All in all, the study shows that machine learning can serve as a powerful and interpretable predictive and optimization framework in predicting and optimizing foam stability, which can be used as a strong tool in data-driven formulation design and in enhanced oil recovery applications.