Machine learning enhanced process design in protein a chromatography

Quality by Digital Design (QbDD) employs in-silico experimentation to reduce wet-lab reliance and accelerate development. Design space identification is critical for QbDD to overcome bottlenecks and streamline process design. Traditional design space identification methods require costly, high-fidelity models. This work introduces a machine learning-enhanced design space identification approach that utilises wet-lab data. A transfer learning framework is developed under limited data availability by leveraging synthetic datasets created using off-the-shelf mechanistic models. An artificial neural network is constructed and used to classify feasible design regions under high, moderate, and low data availability (HDA, MDA, LDA). Results show strong performance in HDA for the data-driven method, with transfer learning improving accuracy in MDA and being essential in LDA. This approach demonstrates machine learning's potential to enable cost- and time-efficient process design in early-stage biopharmaceutical development.