Implementation of real-time incremental learning for ensemble hybrid model prediction in pilot scale bubble column aeration

Digital twins are frequently discussed in a bio-manufacturing context, but actual realisations of digital twins are rare. To use digital twin instances, significant investments in digital infrastructure and high-fidelity mathematical models are required. This work presents a real-time implementation of an ensemble hybrid model with incremental learning for predicting dissolved oxygen concentration in a pilot-scale bubble column. A bootstrap-aggregated hybrid modelling framework is applied for constructing an ensemble of 1000 hybrid models using different partitions of the training/validation data, providing a measure of the parameter distributions and prediction uncertainty. Each model in the ensemble hybrid model has the same model structure relying on first-principles material balances and an Artificial Neural Network for prediction of the liquid phase volumetric mass transfer coefficient. Incremental learning is applied, efficiently enabling the model to adapt to new data acquired during runtime. The software implementation follows recent ISO issues using a modular structure allowing for flexible allocation of server resources and an intuitive User-Interface is developed for controlling the application. From a real-time prediction study, the models using incremental learning are found to have superior performance both at normal operating conditions, when interpolating, and when extrapolating compared to using only the pre-trained model. • Implementation of a digital twin for pilot-scale bubble column. • Hybrid semi-parametric model combining material balances and an artificial neural network. • Incremental learning enabling model assimilation of new process data during runtime. • Uncertainty Quantification using bootstrapping. • Real-time model predictions and control of application through intuitive User-Interface.