Enabling Bioprocess Upscaling Prediction Through Hybrid Modelling and Transfer Learning Under Small‐Data Scenarios

Accurate prediction of bioprocess scale-up is critical for accelerating the deployment of novel and sustainable biomanufacturing systems. However, this remains challenging as multi-scale data is expensive to generate and mechanistic understanding is often incomplete, leading upscaling decisions to rely heavily on empirical expertise. This work proposes a data-efficient strategy that integrates hybrid modelling with transfer learning to construct a high-fidelity model from limited lab scale data and then adapt it using only pilot scale information for industrial scale prediction. A key innovation is the explicit assessment of this framework under small-data scenarios, reflecting the practical constraints of industrial development. Using a real-world yeast fermentation case, the proposed framework achieved accurate industrial scale dynamic predictions with a mean absolute percentage error of 23.2%, demonstrating its high data efficiency. Furthermore, this study reveals that the greyness of hybrid models exerts a decisive influence on its predictive accuracy and the feasibility of transfer learning under data scarcity, and that its optimal level differs from scenarios with abundant data. These findings therefore provide the first guidance on how to exploit hybrid modelling and transfer learning to build scalable digital twins when data are limited, enabling more confident and reliable bioprocess development and upscaling.