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

Abstract

Surface-enhanced Raman scattering (SERS) spectroscopy is still considered poorly reproducible despite its numerous advantages and is not a sufficiently robust analytical technique for routine implementation outside of academia. In this article, we present a self-supervised deep learning-based information fusion technique to minimize the variance in the SERS measurements of multiple laboratories for the same target analyte. In particular, a variation minimization model, coined the minimum-variance network (MVNet), is designed. Moreover, a linear regression model is trained using the output of the proposed MVNet. The proposed model showed improved performance in predicting the concentration of the unseen target analyte. The linear regression model trained on the output of the proposed model was evaluated by several well-known metrics, such as root mean square error of prediction (RMSEP), BIAS, standard error of prediction (SEP), and coefficient of determination (R²). The leave-one-lab-out cross-validation (LOLABO-CV) results indicate that the MVNet also minimizes the variance of completely unseen laboratory datasets while improving the reproducibility and linear fit of the regression model. The Python implementation of MVNet and the code for the analysis can be found on the GitHub page https://github.com/psychemistz/MVNet.