Physics‐encoded transfer learning for scale‐up modeling of CHO cell bioreactors

Abstract Developing reliable predictive models for mammalian cell bioreactors, particularly Chinese hamster ovary (CHO) cultures widely used in biopharmaceutical manufacturing, remains challenging due to severe data scarcity in industrial‐scale reactors. While laboratory‐scale experiments generate rich data sets, large‐scale runs are expensive and sparsely sampled, making data‐driven modeling unreliable. Transfer learning (TL) from laboratory to production scale therefore offers a promising strategy for scale‐up modeling. However, conventional TL approaches based on freezing and fine‐tuning perform poorly because scale‐induced shifts in transport phenomena and hydrodynamics alter the effective input‐output relationships of the process. In this work, we formulate bioprocess scale‐up as a structured transfer learning problem where biological dynamics remain largely invariant while transport environments introduce scale‐dependent modulation. We propose a physics‐encoded TL method using Feature‐wise Linear Modulation (FiLM) for data‐efficient cross‐scale modeling, achieving improved prediction accuracy compared with conventional TL and direct training on large‐scale data.