Accelerating bioprocess digital twin development by integrating hybrid modelling with transfer learning
Luca Riezzo, Harry Kay, Yansong Feng, Keju Jing, Dongda Zhang
Abstract
• Hybrid modelling and transfer learning accelerate bioprocess digital twin design. • Transfer learning adapt hybrid model to new strains, enhancing predictive accuracy. • Combining AI with kinetic models overcome data scarcity in bioprocess modelling. • Proposed framework generalise to other complex (bio)chemical reaction systems. Biomanufacturing is crucial for sustainable production, yet challenges often arise in industrialising new bioprocesses, such as the need for novel strains and optimal operating conditions. To accelerate high-fidelity digital twin development for new bioprocess design, this study explores integrating hybrid modelling and transfer learning, combining mechanistic models with machine learning to enhance accuracy and predictive performance. To demonstrate the proposed strategy, microalgal production of lutein, a valued carotenoid in pharmaceuticals and food industries, was used as a case study. Initially, a hybrid model was developed to simulate process dynamics of an algal strain Chlorella Sorokiniana XMU17 (source domain) for lutein synthesis. Transfer learning was then employed to adapt the hybrid model to a newly engineered nanoparticle-enhanced strain, Chlorella Sorokiniana F31 , using minimal experimental data. Experimental validation confirmed the high predictive accuracy of the hybrid model, which exhibited significantly lower uncertainties compared to a kinetic model. Most importantly, transfer learning effectively addressed data scarcity challenges for new bioprocess predictive modelling. This study highlights the potential of combining advanced machine learning techniques and physical knowledge to revolutionise novel bioprocess design, accelerate strain performance evaluation, and reduce the time and cost of collecting experimental data.