Accurate modelling of drying kinetics is required to design and control solar drying systems, but the semi-empirical equations used for this purpose depend only on time and are calibrated under nominally steady conditions, which limits their applicability when the drying air conditions vary during the day. This study evaluates whether neural network models that receive the instantaneous drying conditions can predict the drying kinetics of beef in a tunnel-type solar dryer more accurately than the Newton, Page, and Midilli et al. equations. Four architectures (MLP, LSTM, GRU, and TCN) and hybrid versions of them, in which a neural network modulates the drying constant of a Page-type equation, were optimized with four search strategies (an evolutionary algorithm, Bayesian optimization, random search and Hyperband) under an identical budget of 300 configurations per run and five independent runs per strategy. The models were trained and validated with four drying series measured on two days, using batch-level cross-validation, and evaluated once on four tray-level series collected on two other drying days. Bayesian optimization reached the lowest validation error for seven of the eight architectures ( p<0.05 , Kruskal–Wallis over independent runs), while Hyperband required the least training time for the plain LSTM, GRU, and TCN and for the hybrid TCN. Replacing the ambient air conditions with the conditions measured next to the tray reduced the mean test error by 21 p<0.05 ), and the hybrid formulation guaranteed a monotonic curve with an interpretable drying constant.
Carbon quantum dots (CQDs) have been widely studied as sensors for chemical hazards in food. Lower limits of detection (LODs), more elaborate sensor designs, and tests in real food are often taken as evidence of progress. Yet the conditions required for comparing LODs across studies, or for tracing changes in sensor design, have received less attention. We conducted a critical longitudinal evidence synthesis of 106 original studies published from 2016 to 2026. Each study was coded on three independent axes: material architecture, recognition architecture, and signal/readout architecture. We then examined target-specific LOD patterns and separated real-food testing from stronger validation evidence. Only tetracycline and Hg(II) formed repeated target-specific series large enough for the main temporal analysis. Reported LODs tended to decrease with publication year among tetracycline fluorescence studies, but the bootstrap interval included zero and later studies more often used ratiometric readouts. No comparable temporal association was found for Hg(II). Material and recognition architectures did not move steadily from simple systems toward composite designs. Descriptively, the clearest directional change occurred in the readout layer, where ratiometric, dual-emission, and multimodal systems became more common. Real-food testing and multi-matrix evaluation also appeared more often in later studies. Reference-method comparisons and evidence from native-positive, incurred, or food certified reference materials did not show similar growth. Portable formats increased only modestly. These findings indicate uneven development across design, target-specific sensitivity, food-matrix testing, validation evidence, and deployment. LOD alone is therefore an incomplete measure of practical value.
Nanguo pear easily generates latent bruises from mechanical collision during postharvest handling. Early latent bruises show faint characteristic signals, and single-spectrum or single-texture detection fails to deliver high precision due to overlapping feature spaces. This study develops a hyperspectral detection model integrating spatial-spectral nonlinear fusion (SSNF) and Stacking ensemble learning to quantify the earliest detectable time limit of latent bruises. Uninformative variable elimination (UVE) screens core characteristic wavelengths, while principal component analysis (PCA) extracts tissue-damage-sensitive texture features. A second-order multiplicative coupling matrix is built to model nonlinear interactions between spectral and texture data. Random forest (RF), LightGBM, and SVM serve as heterogeneous base learners, with logistic regression (LR) acting as the meta-learner for two-layer decision fusion. The results demonstrated that the fusion model achieved an average recall of 98.2
The overuse of artificial sweeteners (ASs) in food and beverages would induce chronic diseases for human. In this paper, a new high-performance liquid chromatography (HPLC) method was developed to simultaneously evaluate and quantify five common ASs (acesulfame-K, saccharinnatrium, aspartame, alitame, and neotame). The physicochemical property and molecular structure of ASs were calculated by ACD/Precepta and density functional theory (DFT); these parameters were used to assist the development of the HPLC method. Following the optimization of chromatographic parameters, including column, mobile phase, and column temperature, the detection range for ASs was established between 1 and 100 mg/L; all analytical curves showed R2 ≥ 0.999 with LOD were 0.1 or 0.5 mg/L. This method was effectively applied to quantify five ASs in various custom-blended beverages, demonstrating satisfactory detection performance that complies with the standards of the Codex Alimentarius set by the FAO/WHO and the GB 2760-2024 regulations established by the Chinese government.
Accurate and rapid determination of both fat and protein in meat is crucial for nutritional labelling, quality grading, and regulatory compliance. Near-infrared spectroscopy (NIRS) enables fast, non-destructive measurement, yet its multivariate calibration is hampered by severe collinearity and matrix-induced spectral interference. This study presents an incremental-improvement hybrid wavelength selection workflow, S-NAS-SPA, which integrates a supervised net analyte signal (S-NAS) filter with the successive projections algorithm (SPA). The S-NAS stage builds, for each wavelength, a PCA-truncated interference subspace (q set to 10, which captures 100
