Physics-Guided Residual Learning with Conditional Modulation for Quality Monitoring of Kiwifruit Juice During...

Area of Science:

• Food Science
• Chemical Engineering
• Data Science

Background:

• Predicting juice quality during pasteurization is difficult due to complex kinetics and scarce labeled data.
• Existing methods struggle with the interplay of temperature, time, and degradation reactions.

Purpose of the Study:

• To develop a novel physics-guided residual learning framework for dynamic quality monitoring during juice pasteurization.
• To integrate near-infrared spectroscopy and electronic nose data for enhanced prediction accuracy.
• To address challenges posed by limited labeled data in controlled pasteurization settings.

Main Methods:

• A physics-guided residual learning (PIRL) framework incorporating a first-order kinetic equation as a physical prior.
• A Feature-wise Linear Modulation-inspired conditional modulation (FiLM-ICM) mechanism to adaptively rescale spectral and olfactory features based on temperature and time.
• Fusion of near-infrared spectroscopy and electronic nose signals for data-driven compensation of unmodeled deviations.

Main Results:

• The PIRL-FiLM framework demonstrated superior performance over conventional chemometric baselines in predicting quality indicators.
• Achieved R-squared improvements up to 0.030 and RPD gains exceeding 0.230 on a kiwifruit juice dataset.
• Validated through leave-one-trajectory-out cross-validation, confirming robustness.

Conclusions:

• Physics-guided learning and data-driven approaches are complementary, not competing, for robust quality monitoring.
• The hybrid strategy offers a practical solution for process analytical technology, especially under small-sample constraints.
• This framework enables accurate quality monitoring in laboratory-scale development and small-batch production where large datasets are unavailable.