Unsupervised optimization of spectral pre-processing selection to achieve transfer of Raman calibration models

• Pre-processing optimization is proposed as an unsupervised model transfer method. • Maximum Mean Discrepancy (MMD) used as metric for pre-processing selection. • MMD quantifies the distance between predictions in source and target domains. • Method demonstrated on simulations, industrial use cases and benchmarked. • Consistent target domain generalization achieved for covariate shift scenarios. Spectral pre-processing is a crucial step in the development of calibration models for spectroscopic process analytical technology (PAT). Typically, pre-processing is optimized in the source domain, which can hinder deployment to new (target) domains as spectral differences may arise when sample or instrument changes occur. While specialized methods for calibration transfer and maintenance exist, this work suggests using pre-processing alone as a straightforward method to achieve model transfer. Maximum mean discrepancy (MMD) is used as an unsupervised metric for selecting a model (pre-processing) with optimized performance in the target domain. MMD utilizes unlabeled target spectra to quantify the distance between the distributions of the source and target predictions of qualified candidate models (models with acceptable source domain performance) and the model with minimal discrepancy between the predictions is selected. Applicability of the MMD-based pre-processing selection was initially explored on simulations of different types of dataset shifts (covariate, conditional and prior), which suggested that the method achieved successful transfer primarily under covariate shift. Next, the approach was applied to two use cases from the biopharmaceutical industry (varying process and acquisition settings; varying instruments from different manufacturers), where the MMD method achieved lower error of prediction in the target domain compared to a cross-validation based pre-processing selection. Finally, the proposed approach was benchmarked with domain-invariant partial least squares showing superior performance of pre-processing while highlighting benefits of combining both approaches. Overall, this work presents a novel and easy-to-adopt unsupervised model transfer method applicable to many common transfer scenarios encountered in bioprocessing.