Abstract
Raman spectroscopy is increasingly used as a process analytical technology (PAT) for real-time monitoring of bioprocesses. However, chemometric models developed using high-throughput (HT) mini-bioreactor systems often show limited predictive performance when applied to larger-scale processes, reflecting an out-of-distribution (OOD) challenge in cross-scale model transfer. In this study, we investigated whether variable-specific data preprocessing strategies can improve the cross-scale prediction performance of Raman chemometric models calibrated using HT cell culture data. Multiple preprocessing approaches were systematically evaluated for key cell culture target outputs, and the optimal preprocessing pipeline for each output was selected based on its ability to minimize cross-scale prediction error. The optimized variable-specific preprocessing pipelines reduced cross-scale prediction errors by 14.0-56.1% in RMSEP across the five initially poorly transferable target outputs compared with the standard single-scale preprocessing pipeline. Consequently, these optimized single-scale models achieved predictive accuracy approaching that of resource-intensive combined-scale models. Importantly, the improvement was not dependent on a specific preprocessing technique but rather on selecting preprocessing strategies suited to the spectral characteristics of each target output. These findings suggest that intelligent preprocessing selection can mitigate out-of-distribution effects in Raman chemometric models and enable more reliable prediction across scales using HT datasets for model calibration. This approach highlights the potential of Raman spectroscopy as a practical and scalable PAT tool for industrial bioprocess development.