Self-supervised learning for inter-laboratory variation minimization in surface-enhanced Raman scattering...

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Surface-Enhanced Raman Scattering (SERS) spectroscopy offers significant advantages but suffers from poor reproducibility, limiting its routine analytical application.
• Variability in SERS measurements across different laboratories hinders its adoption outside academic research.

Purpose of the Study:

• To develop a self-supervised deep learning technique for information fusion to minimize variance in multi-laboratory SERS measurements.
• To enhance the robustness and reproducibility of SERS spectroscopy for reliable quantitative analysis.

Main Methods:

• A novel variation minimization model, termed the minimum-variance network (MVNet), was designed using a self-supervised deep learning approach.
• A linear regression model was trained on the output of MVNet to predict analyte concentration.
• Leave-one-lab-out cross-validation (LOLABO-CV) was employed to assess model performance on unseen laboratory data.

Main Results:

• The MVNet demonstrated improved performance in predicting the concentration of unseen target analytes.
• The regression model trained with MVNet output showed enhanced metrics including RMSEP, BIAS, SEP, and R-squared.
• LOLABO-CV confirmed that MVNet minimizes variance in datasets from completely unseen laboratories, improving reproducibility and linear fit.

Conclusions:

• The proposed MVNet effectively minimizes inter-laboratory variance in SERS measurements, significantly improving reproducibility.
• This deep learning approach enhances the robustness of SERS spectroscopy, making it more suitable for routine analytical implementation.
• The study provides a validated computational framework and open-source code for advancing SERS data analysis.