Gas Holdup Prediction Using Machine Learning Approaches

Gas holdup estimation is important in gas-liquid systems while designing chemical, petroleum, and environmental engineering processes. The aim of this project was to develop an XGBoost algorithm-based machine learning model to forecast gas holdup from very simple to moderate characteristics varying at the start; these are superficial gas velocity, sparger hole diameter, gas and liquid densities, liquid height, additive and ion concentration ratio, and concentrations of ions and additives. Preprocessing of data involved the step of shuffling between 4043 data points and normalizing them to create consistency. Lasso Regression, Recursive Feature Elimination, and feature importance analysis aided in reducing the best predictors, making the model simpler without affecting accuracy. Five machine learning algorithms—Decision Trees (DT), Random Forest (RF), Support Vector Machines (SVM), Gradient Boosting (GB), and XGBoost—were compared on R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, MAE, and RMSE. The R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> plot shows that XGBoost outperformed all the models, with 0.9595 accuracy using 12 features and 0.9545 using 7 features. Gradient Boosting came close, and Random Forest showed competitive but inconsistent accuracy. SVM was consistent but lower in performance, and Decision Trees showed the lowest R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values, indicating poor generalization. An interactive Shiny app was developed to apply a truncated XGBoost model for real-time predictions with unnormalized inputs. The project demonstrates XGBoost's potential and machine learning's potential for accurate gas holdup prediction in industrial processes.