Advanced modeling techniques for predicting gas holdup in bubble columns using machine learning

Bubble columns are critical in bioreactor operations for biofuel production and microbial carbon sequestration, influencing hydrodynamics and gas-liquid mass transfer. This study delves into machine learning regression models for gas hold up prediction in bubble columns including linear regression, support vector regressor, decision trees, gaussian process regressor, random forest, artificial neural networks, XGBoost, and ANN ensembles. The model selection process carefully addressed the bias-variance trade-off, evaluated data adequacy through learning curves, assessed hyperparameter tuning via validation curves, applied nested cross-validation to ensure robust generalizability and included deployment of the models on new unseen data. XGBoost achieved the best test performance (MSE = 0.0004, R2 = 0.97, MAE = 0.014), while Linear Regression with feature engineering performed comparably on unseen data. These findings propose a systematic model selection procedure using ML in regression tasks, providing a resource-efficient alternative to CFD methods for modeling gas holdup dynamics as medium properties evolve.