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

Area of Science:

• Spectroscopy
• Chemometrics
• Machine Learning

Background:

• Deep learning in Raman spectroscopy allows rapid analysis but faces challenges with model portability due to interdevice variations.
• Systematic differences between spectrometers hinder the direct application of trained models across different instruments.

Purpose of the Study:

• To introduce a parameter-efficient method for calibration transfer in deep learning-based Raman spectroscopy.
• To enable seamless model portability across diverse spectrometers with minimal data and computational resources.

Main Methods:

• Proposed a Low-Rank Adaptation-based Calibration Transfer (LoRA-CT) method for parameter-efficient fine-tuning.
• Decomposed weight updates into low-rank matrices to reduce trainable parameters significantly (600× reduction vs. full fine-tuning).
• Validated LoRA-CT across three datasets including solvent mixtures and blended oils.

Main Results:

• LoRA-CT demonstrated superior calibration transfer performance using very few transfer samples compared to conventional methods.
• Achieved R² = 0.952 and RMSE = 0.072 on the methanol mixture test set, outperforming piecewise direct standardization and full parameter fine-tuning.
• Experimental results confirmed significant improvements in accuracy and efficiency.

Conclusions:

• LoRA-CT establishes a new paradigm for resource-efficient spectroscopic model deployment.
• The modular, plug-and-play design facilitates dynamic switching between spectrometers.
• This method is particularly advantageous for portable spectrometers and multi-instrument systems facing sample scarcity and computational constraints.