Hybrid AI modeling techniques for pilot scale bubble column aeration: A comparative study

Abstract

With increased accessibility of process data from the production lines in chemical and biochemical production plants, the use of data-based modeling methods is gaining interest. In this work, three different data-based modeling approaches are applied for modeling aeration in a pilot scale bubble column. In all three modeling approaches the same serial hybrid-model structure is used, combining a species conservation balance based on first-principles with different data-based models for the overall volumetric mass transfer coefficient ($K_{L}a$). Simple empirical correlations with parameters fit to process data provide transparent models but lack the accuracy of Artificial Neural Networks (ANNs). ANNs provide models with high accuracy within the operation regimes used for training, however, the models are prone to overfitting, and their black-box nature results in models that are difficult to interpret. As an alternative, a symbolic regression-inspired technique is used for discovering symbolic equations, resulting in interpretable models with accuracy that is comparable to that of the ANN.