Predictive Modeling of Mean Residence Time in Bubble Column Reactors: A Machine Learning Approach Using Linear Regression, Random Forest, and Neural Networks

ABSTRACT: Accurately predicting Mean Residence Time (MRT) in bubble column reactors plays a pivotal role in enhancing industrial chemical processes. This research presents a machine learning approach to predict MRT using three models: Linear Regression, Random Forest, and Neural Networks (MLP Regressor). The dataset is drawn from an Extended Three-Phase Fluidized Bed Reactor RTD system, incorporating key variables such as gas flow rate, liquid viscosity, particle density, and reactor dimensions. To assess the models' performance, metrics like Mean Absolute Error (MAE), Mean Squared Error (MSE), and R² score were employed. The experimental analysis indicated that the Neural Network model outperformed both Linear Regression and Random Forest, effectively capturing the complex non-linear interactions within the dataset. The Neural Network delivered lower MAE and MSE values, along with a more accurate fit to the data, showing a slightly better R² score. However, both Random Forest and Linear Regression models underperformed, with the Random Forest model lagging due to insufficient hyperparameter tuning. Additionally, key factors affecting MRT, such as higher void fractions and lower viscosities, were identified as contributors to extended residence time. These results highlight the critical role of model tuning and feature engineering in refining predictions. Although the Neural Network displayed higher accuracy, further optimization and feature extraction are necessary to improve overall performance. Cross-validation, outlier detection, and advanced models like Gradient Boosting Regressor or XGBoost are recommended for future research. This method presents a promising direction for enhancing MRT predictions in bubble column reactors and optimizing chemical processes.