A differential evolution-based joint optimization method for full-process near-infrared spectral modeling and its...

Area of Science:

• Analytical Chemistry
• Chemometrics
• Spectroscopy

Background:

• Conventional near-infrared (NIR) spectral modeling involves fragmented, stepwise optimization.
• This traditional approach relies heavily on manual expertise and fails to capture interactive effects.
• Limitations lead to suboptimal predictive performance and poor generalization in NIR models.

Purpose of the Study:

• To develop a unified, automated optimization method for NIR spectral modeling.
• To address the limitations of stepwise optimization in conventional NIR analysis.
• To enhance the predictive accuracy, robustness, and generalization of NIR models.

Main Methods:

• Proposed a differential evolution (DE)-based full-process joint optimization method.
• Integrated outlier removal, spectral preprocessing, wavelength selection, and regression algorithm optimization.
• Employed a cross-validation framework to minimize root mean square error (RMSE) and an ensemble strategy for robustness.

Main Results:

• Achieved high prediction accuracy using a minimal number of selected wavelengths (≤10).
• Demonstrated consistent generalization across diverse datasets and instrument platforms.
• Exhibited chemical interpretability, validating the selected features.

Conclusions:

• The DE-based joint optimization method offers a practical approach for NIR spectral modeling.
• Enables the development of low-cost, task-specific, and lightweight NIR sensors.
• Advances NIR analysis towards intelligent and application-oriented deployment.