Advancing near-infrared spectroscopy: A synergistic approach through Bayesian optimization and model stacking.

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Chemometrics

Background:

• Fourier transform near-infrared (FT-NIR) spectroscopy is a key non-destructive analytical technique.
• Portable FT-NIR systems enhance accessibility, but model optimization remains crucial.
• Partial Least Squares Regression (PLSR) is a standard chemometric method for compositional analysis.

Purpose of the Study:

• To develop an optimized PLSR modeling framework for FT-NIR spectroscopy.
• To enhance accuracy, adaptability, and automation in material analysis.
• To improve upon traditional grid search methods for PLSR model optimization.

Main Methods:

• A novel framework combining Bayesian search and model stacking was implemented.
• Bayesian search efficiently explored the hyperparameter space for PLSR models.
• Stacked models integrated knowledge from multiple optimized PLSR models.

Main Results:

• Achieved a 51.5% reduction in training RMSE and 26.1% in testing RMSE.
• Increased R-squared by 10.9% (training) and 10.4% (testing).
• Reduced model optimization time by approximately 90% compared to grid search.
• Demonstrated improved robustness against instrumental variations, with a 24.1% reduction in prediction mean range.

Conclusions:

• The Bayesian-stacked approach significantly enhances PLSR model performance in FT-NIR spectroscopy.
• This method provides a more efficient, automated, and robust solution for material analysis.
• The findings support broader adoption of advanced chemometric techniques in spectroscopic applications.