Calibration Transfer of Deep Learning Models among Multiple Raman Spectrometers via Low-Rank Adaptation.

While deep learning-enhanced Raman spectroscopy enables rapid sample analysis, model portability among spectrometers remains hindered by systematic interdevice variations. In this study, a Low-Rank Adaptation-based Calibration Transfer method (LoRA-CT) is proposed to perform parameter-efficient fine-tuning of deep learning models across spectrometers. By decomposing weight updates into low-rank matrices, LoRA-CT achieves superior calibration transfer with minimal samples, reducing trainable parameters by 600× compared to full parameter fine-tuning. Experimental validation across three data sets (solvent mixtures and blended oils) demonstrates that LoRA-CT with very few transfer samples significantly outperforms conventional methods. On the methanol mixture test set, it achieves = 0.952 (vs 0.846 for piecewise direct standardization and 0.863 for full parameter fine-tuning) with lower RMSE (0.072 vs 0.117 and 0.114). The modular design enables dynamic switching between spectrometers through plug-and-play LoRA modules. This work establishes a new paradigm for resource-efficient spectroscopic model deployment, particularly advantageous for portable spectrometers and multi-instrument industrial systems where sample scarcity and computational constraints coexist.