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

Accurate prediction of quality indicators during juice pasteurization is challenged by the intricate coupling between temperature-dependent reaction kinetics and the limited availability of labeled data under laboratory-controlled pasteurization conditions. Here, we present a physics-guided residual learning framework that fuses near-infrared spectroscopy with electronic nose signals for dynamic quality monitoring. A first-order volatilization-saturation kinetic equation is embedded as a physical prior (PIRL), anchoring the prediction to known degradation behavior while reserving the residual learner for data-driven compensation of unmodeled deviations. To further incorporate process control information, a Feature-wise Linear Modulation-inspired conditional modulation (FiLM-ICM) mechanism is deployed, enabling temperature and time to act not as passive covariates but as active modulators that adaptively rescale spectral and olfactory feature responses. Validated on a kiwifruit juice pasteurization dataset under leave-one-trajectory-out cross-validation, the PIRL-FiLM framework consistently outperforms conventional chemometric baselines, achieving R2 improvements up to 0.030 and RPD gains exceeding 0.230. This hybrid strategy demonstrates that physical knowledge and data-driven learning are not competing but complementary, offering a robust and interpretable paradigm for process analytical technology under small-sample constraints. This framework offers a practical solution for quality monitoring in scenarios where large-scale labeled datasets are unavailable, such as in laboratory-scale process development and small-batch production.