On the estimation of sugars concentrations using Raman spectroscopy and artificial neural networks

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Chemometrics
• Machine Learning

Background:

• Accurate quantification of sugars (glucose, sucrose, fructose) is crucial in food science and quality control.
• Traditional methods can be time-consuming or require extensive sample preparation.
• Spectroscopic techniques offer rapid, non-destructive analysis possibilities.

Purpose of the Study:

• To evaluate the performance of Raman spectroscopy coupled with feed-forward neural networks (FFNN) for sugar concentration estimation.
• To compare FFNN performance against other chemometric methods (SVM, LDA, LR, iPLS).
• To assess the applicability of this combined method for both aqueous solutions and solid food matrices.

Main Methods:

• Raman spectra were acquired for glucose, sucrose, and fructose in water solutions.
• Feed-forward neural networks (FFNN) were employed for classification and non-linear fitting to estimate concentrations.
• Performance was benchmarked against Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), Linear Regression (LR), and interval Partial Least Squares (iPLS).

Main Results:

• FFNN achieved a best-case classification accuracy of 93.33% and a Root Mean Square Error of Prediction (RMSEP) of 3.51% for water solutions.
• FFNN significantly outperformed LDA (82.22% classification), SVM, LR, and iPLS in estimating sugar concentrations.
• For solid food products (donuts, cereal, cookies), the FFNN method achieved an RMSEP of 1%.

Conclusions:

• Raman spectroscopy combined with FFNN provides a highly accurate and efficient method for quantifying sugar concentrations.
• The FFNN approach demonstrates superior performance compared to traditional chemometric models for both solutions and complex food matrices.
• This technique holds significant potential for rapid quality control and analysis in the food industry.