Tailored Deep Learning-Assisted In Situ SERS: Overcoming Surface Irregularities-Induced Large Signal Variation on...

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Machine Learning

Background:

• Surface-enhanced Raman spectroscopy (SERS) faces challenges in accurate quantification on uneven biological surfaces due to signal variability.
• Surface irregularities and the coffee-ring effect significantly impact SERS data reproducibility.

Purpose of the Study:

• To develop a deep learning-assisted in situ SERS strategy for highly reproducible quantification on uneven biological surfaces.
• To address and mitigate spectral variability in SERS measurements of biological samples.

Main Methods:

• A tailored one-dimensional convolutional neural network (1D-CNN) with decreasing kernel sizes was developed.
• Minimal sample preparation and a low-cost SERS substrate were employed.
• The 1D-CNN model was compared against traditional methods like single-peak intensity calibration (SPIC) and random forest.

Main Results:

• The tailored 1D-CNN significantly improved the quantification of thiabendazole on apple skin, increasing R² from 0.332 (SPIC) to 0.935.
• The deep learning approach demonstrated superior performance over other calibration models.
• The model achieved a short training time of 242 seconds.

Conclusions:

• A deep learning-enabled framework effectively mitigates surface-induced signal variability in in situ SERS.
• This strategy enhances quantitative robustness for analyzing heterogeneous biological samples.
• The approach offers a promising solution for rapid surface analysis and residue screening.