[LLE-PLS nonlinear modeling method for near infrared spectroscopy and its application]

Area of Science:

• Analytical Chemistry
• Chemometrics
• Spectroscopy

Context:

• Traditional Near Infrared (NIR) spectra modeling using Partial Least Squares (PLS) struggles with nonlinear correlations.
• Locally Linear Embedding (LLE) is a novel manifold learning algorithm for nonlinear dimension reduction.
• No prior application of LLE in NIR spectra information processing has been reported.

Purpose:

• To propose and evaluate a novel nonlinear modeling method, LLE-PLS, for NIR spectra analysis.
• To combine LLE for dimension reduction and PLS for regression in NIR spectral data.
• To assess the performance of LLE-PLS in correlating NIR spectra with salvia acid B concentrations.

Summary:

• The LLE-PLS method was applied to predict salvia acid B concentrations in column chromatography eluates.
• LLE-PLS demonstrated superior performance compared to conventional preprocessing methods like MSC, derivatives, and normalizations.
• Optimized LLE-PLS achieved high accuracy with a minimum RMSECV of 0.128 mg/mL and R² of 0.9988, outperforming standard PLS.

Impact:

• LLE-PLS effectively models nonlinear correlations between NIR spectra and physicochemical properties.
• The method enables accurate prediction and shows potential for online monitoring of processes like column chromatography.
• Parameter selection (nearest neighbor k and output dimension d) influences LLE-PLS performance, with k being robust and d requiring careful optimization.