Model-Driven Scale-Up from High-Throughput to Bench-Top Scale

Bioprocess development for microbial systems typically requires extensive experimentation across multiple scales, from high-throughput miniaturized systems to lab-scale bioreactors. This study presents a hybrid modeling framework that integrates data from miniaturized high-throughput microbioreactor experiments (BioLector XT, ~800 µL) and 5 L lab-scale fed-batch cultivations of Escherichia coli ATCC 25922GFP for the expression of GFP under IPTG induction.

A model-based design of experiments (MBDoE) approach was employed to systematically vary key process parameters, induction time, inducer (IPTG) concentration, feed rate, and temperature, generating 96 screening runs across 24 conditions and 10 large-scale runs across 6 conditions. The hybrid model, combining mechanistic mass balances with Gaussian process machine learning for the specific growth and production rates, was trained using a bootstrap aggregation strategy to improve robustness and quantify prediction uncertainty. Cross-scale normalization was applied to reconcile sensor differences between scales prior to joint modeling. The resulting model achieved test set relative RMSE values of approximately 0.19–0.20 for both biomass and GFP, comparable to the intrinsic biological replicate variability.

A systematic transfer learning and run-to-run optimization analysis further demonstrated that models informed by small-scale data require as few as one large-scale calibration run to achieve strong predictive accuracy at scale. These results highlight the potential of hybrid modeling to substantially reduce the number of costly large-scale experiments required during bioprocess development and scale-up.