Integrating physical knowledge with machine learning is critical to developing industrially-focused digital twins for monitoring and optimisation of biomanufacturing systems. However, identifying the correct model structure to quantify kinetic mechanisms poses a challenge for the construction of mechanistic and data-driven models. This study proposes a hybrid modelling strategy comprising of a simple kinetic model to describe the overall process trajectory and a data-driven model to estimate the mismatch between the kinetic equations and real process. An automatic model structure identification algorithm is used to identify the most probable kinetic model structure and minimum number of data-driven model parameters that can well represent different bioprocess behaviours over broad operating conditions. Through this approach, a hybrid model was constructed to simulate biomass growth, nutrient consumption, and product synthesis in an algal photo-production process. Performance of this model for predictive modelling, optimisation, and online self-calibration is demonstrated, indicating its advantages for industrial application.

Machine learning techniques have been successfully used to simulate and optimise bioprocesses. This study explores the feasibility to apply Gradient Boosting, an emerging Ensemble Learning algorithm, which combines weak learners to generate better predictions for bioprocess dynamic modelling and prediction. A thorough procedure was presented for Gradient Boosting based data-driven model construction. Different case studies were employed including fermentation and algal photo-production processes. Given that generating a large size of experimental data for model training is time consuming and challenging to many bioprocesses, this work launched a first investigation on the data efficiency of Gradient Boosting by comparing its predictive capability against the predominantly used artificial neural networks. By carrying out a series of experimental verifications over a broad spectrum of process operating conditions, this study concluded that Gradient Boosting may have several advantages in small experimental datasets and can outperform artificial neural networks for bioprocess predictive modelling, indicating its potential for future bioprocess digitalisation and optimisation.