Randomized trial investigates unsupervised transfer learning for analyte prediction in bioprocess settings, suggesting enhancements in monitoring strategies.
Key Points
The aim is to evaluate whether unsupervised transfer learning methods can improve model generalization for analyte prediction from Raman spectra across different bioprocess conditions.
Assessed four unsupervised transfer learning methods: CORAL, JDOT, TCA, and OPP-MMD.
Benchmarking performed against two baselines: a fixed preprocessing workflow and an optimized source calibration.
Evaluation conducted on two Raman datasets focusing on glucose, biomass, and phosphate predictions.
JDOT and CORAL improved predictions significantly under high cross-domain variability.
No benefit was observed in scenarios with high within-domain variability; performance was lower than the baseline in those cases.
Unsupervised transfer learning methods can be integrated effectively into standard Raman preprocessing workflows.