Chinese Baijiu is a kind of complicated composition distilled liquor, its quality and authenticity mainly depend on its alcohol by volume (ABV), and many kinds of complex trace fragrant compounds. Samples for traditional analysis methods are often time-consuming and destructive, thus limiting its applicability in industrial rapid screening. In this study, the deep learning algorithms were combined with the oblique-incidence reflectivity difference (OIRD) method, and two hybrid models of TCN-Transformer-KAN and CNN-LSTM-ARIMA were selected to detect the three kinds of simulated liquor flavor samples with added ethyl acetate, ethyl lactate and isoamyl alcohol, along with 22 commercial Baijiu with different flavors and ABVs, respectively. The prediction and classification accuracy of the two models were higher than 96%. The experimental results indicate that OIRD can achieve a rapid, non-destructive, accurate and reliable detection of Baijiu.

Food-grade interfacial materials with different molecular structures often exhibit distinct adsorption behaviors at oil–water interfaces, yet direct comparisons under identical conditions and across multiple scales remain limited. We hypothesized that the structural differences among α-cyclodextrin, sucrose fatty acid ester, and sodium caseinate would result in distinct adsorption behaviors and interfacial organizations that could be revealed across multiple scales. To test this hypothesis, a multiscale characterization strategy combining molecular docking, confocal laser scanning microscopy, oblique-incidence reflectivity difference, and ellipsometry was applied to the soybean oil–water interface. Molecular docking provided molecular-level information on possible interactions between representative structures of the three materials and a triglyceride model molecule. Confocal laser scanning microscopy revealed clear differences in interfacial organization: α-cyclodextrin exhibited the most continuous interfacial distribution, sucrose fatty acid ester showed a connected but less uniform distribution, whereas sodium caseinate displayed heterogeneous and discontinuous interfacial domains. Oblique-incidence reflectivity difference monitoring further revealed distinct time-dependent responses. At sufficient concentrations, α-cyclodextrin exhibited a persistent low-signal state, sucrose fatty acid ester showed a clear concentration-dependent response, whereas sodium caseinate produced a pronounced initial response followed by signal recovery during prolonged monitoring. The optical responses were further matched with the effective interfacial thickness obtained by ellipsometry, and Boltzmann fitting revealed distinct nonlinear optical response–thickness relationships among the three systems. Taken together, these multiscale results support the hypothesis that chemically different food-grade interfacial materials exhibit distinct patterns of interfacial organization and temporal evolution, highlighting the importance of considering multiple interfacial characteristics when comparing their adsorption behaviors at oil–water interfaces.

Different external conditions in the aqueous phase have a significant impact on the adsorption behavior of (3-cyclodextrin ((3-CD) at oil-water interfaces, the issues of real-time monitoring of the micrometer-scale adsorption behavior remain unexplored under non-destructive conditions. In this study, a multi-technique method ranging from data and experimental simulation to actual testing was adopted. By utilizing molecular docking, quartz crystal microbalance with dissipation (QCM-D) and oblique-incidence reflectivity difference (OIRD), the competitive and ionic effects were characterized under three different external conditions (Sodium caseinate, Tween 20 and NaCl of 1 g/100 mL). The multiscale properties of (3-CD adsorption behavior at the oilwater interface detected using the combination of simulation methods and advanced characterization techniques, especially OIRD, are helpful in improving our understanding of the preparation of Pickering emulsions.

Whey protein is widely consumed as a high-quality nutritional supplement throughout the sports- and functional-food industries. Its fraudulent substitution with lower-cost ingredients such as soy protein isolate (SPI) and maltodextrin (MD) compromises product quality, threatens consumer safety, and therefore calls for rapid, label-free, and non-destructive authentication techniques. In this study, oblique-incidence reflectivity difference (OIRD) — an interfacially sensitive optical method that requires no chemical labeling and only seconds of acquisition per measurement — was applied for the first time to the identification of SPI and MD adulteration in whey-protein solutions over a 5–50

The field of protease cleavage site prediction is undergoing a profound transformation driven by the integration of artificial intelligence (AI) technologies within computational bioinformatics frameworks. Proteolysis, a critical mechanism in regulating protein function and cellular processes, involves complex and variable substrate recognition patterns that pose significant challenges to traditional computational approaches. AI can accurately model the heterogeneity of proteolytic mechanisms and improve the generalizability of protease cleavage site prediction. This review examines how AI leverages feature extraction, multidimensional regulatory information integration and nonlinear relationship modeling to address the limited generalizability of traditional methods. Representative studies demonstrate the successful application of AI in cleavage site prediction across different protease families. We identify major obstacles in current AI applications, focusing on data quality, multimodal feature integration and model interpretability. The content systematically outlines the progress, key challenges and future directions in AI-driven protease cleavage site prediction, aiming to provide theoretical guidance for improving prediction efficiency and promoting broader AI applications in the biomedical field.

In the field of cosmetic raw materials, accurately distinguishing plant fermentation extracts from traditional water extracts is crucial for ensuring product quality, user safety, and the authenticity of claimed effects. However, conventional chemometrics has limited capacity to explore the nonlinear characteristics of complex fluorescence spectra, and machine learning still faces shortcomings in feature representation and generalization capability, necessitating the development of more efficient and accurate identification methods. In this study, white mulberry (Morus alba L.) was used as the raw material to prepare 966 samples, including yeast fermentation extracts, lactic acid bacteria fermentation extracts, and water extracts. After collecting the fluorescence spectra of each sample, six deep learning models-Informer, PatchTST, Transformer, TCN, LSTM, and CNN-were constructed, alongside traditional models such as SVM, PCA-LDA, and PLS-DA, to systematically compare classification performance. The experiments showed that the Informer model performed the best, with an accuracy, precision, recall, and F1 score of 0.981, 0.982, 0.981, and 0.981, respectively, surpassing all other deep learning models and significantly exceeding traditional methods such as SVM, PCA-LDA, and PLS-DA, demonstrating superior feature extraction and generalization capabilities. This study integrates fluorescence spectroscopy with deep learning, providing a novel and effective solution for identifying liquid cosmetic raw materials.

Accurately distinguishing between plant fermented extracts and aqueous extracts in cosmetic raw materials is crucial for ensuring product quality, safety, and authentic efficacy, thereby safeguarding consumer interests. In this work, Morus alba L. was used as raw material to prepare 54 samples, including yeast fermentation broth, lactobacillus fermentation broth, and aqueous extract. Based on the fluorescence spectral data of samples, deep learning algorithms were used to establish identification models, which included Informer, TCN, Transformer, LSTM, CNN, and PatchTST. Experimental results showed that the Informer model outperformed the others, achieving high evaluation metrics including precision and recall. The integration of deep learning with fluorescence spectroscopy offers a novel and effective approach for distinguishing liquid cosmetic raw materials.

Honey adulteration detection was of great significance for ensuring food safety. This paper proposed a honey adulteration concentration detection method based on OIRD technology combined with machine learning models. First, the experiment found that there were differences in the raw OIRD time-series optical signals among different honeys with different adulteration types. Secondly, this study preprocessed the raw OIRD time-series optical signals using S-G smoothing and MSC correction, respectively. Then, the time-domain features of the OIRD signals were extracted, and to improve the model's accuracy, the UVE and SPA algorithms were utilized to conduct a secondary screening of the aforementioned features. Finally, taking the screened time-domain features as input, prediction models for honey adulteration concentration were established using Partial Least Squares Regression (PLSR),Support Vector Regression (SVR), Extreme Learning Machine (ELM), and eXtreme Gradient Boosting (XGBoost). The experimental results showed that the XGBoost model constructed using S-G smoothing preprocessing combined with UVE feature screening performed the best. The determination coefficients of the prediction set of this model for the three honey adulteration concentrations were all greater than 0.94, and the root mean square errors of prediction (RMSEP) were all less than 2.2. This study confirmed the feasibility of using OIRD technology for honey adulteration detection and provided a new approach for detecting honey adulteration.

Adulteration identification of olive oil is an essential issue in the field of food-related research. In this work, oblique-incidence reflectivity difference (OIRD) method was used to recognize adulteration edible oils in olive oil. In order to reduce the impact of errors, the real and imaginary signals of OIRD were averaged. For the single edible oil adulterated in olive oil, Transformer model, Sparrow Search Algorithm-Hybrid Kernel Extreme Learning Machine (SSA-ELM) model and extreme gradient boosting (XGBoost) model were used to establish analysis and discrimination model. Experimental suggested that all models exhibited the high prediction accuracy with determination coefficients (R2) of 0.99. Moreover, and Random Forest (RF) model can not only identify the type of adulterated olive oil, but also quantitatively analyze adulterated edible oils in olive oil, with a R2 of 0.98. OIRD method provides a good strategy for solving practical problems in identifying edible oil adulteration.

This study proposes the oblique-incidence reflectivity difference (OIRD) method to overcome the challenges resulting from the complexity of traditional detection methods for the cyclodextrin adsorption process at the water interface. In this paper, OIRD signals were obtained, which contained the information of the dielectric constant and thickness of interface layer formed by three natural cyclodextrins (alpha-, beta-, and gamma-cyclodextrin) lutions at different concentrations. The morphology of the interface film was observed by confocal laser scanning microscopy, and a fitting model between the interface thickness measured by Ellipsometry and the OIRD signal was established. The dissipative particle dynamics method was employed to simulate the adsorption process cyclodextrin particles at the oil-water interface for further analysis and characterization, which verified formation mechanism of interface layer. The experimental results indicate that the OIRD method can achieve non-destructive, accurate and reliable detection of the cyclodextrin adsorption behavior at the oil-water interface.

This study explores the use of near-infrared (NIR), mid-infrared (MIR), and Raman spectral fusion for the rapid prediction of floral origins and main taste components in Apis cerana (A. cerana) honey. Feature-level fusion with the partial least squares regression - random forest (PLSR-RF) model achieved 100 % classification accuracy in identifying floral origins. Additionally, the model demonstrated strong predictive capability for sugars, amino acids, and organic acids, with R2 values ranging from 0.88 to 0.96, and performed exceptionally in predicting total organic acids and amino acids (R2 of 0.94 and 0.93, respectively). The PLSR-RF model showed effective clustering for proline, glucose, and fructose, achieving a 23.5 % improvement in predictive accuracy compared to data-level fusion. These findings confirm the efficacy of the PLSR-RF model for quantitative analysis of A. cerana honey.

Pesticide residue detection plays an important role in vegetable quality and food safety. In this work, we propose a method for detecting difenoconazole pesticide residues based on fluorescence spectroscopy technology and machine learning algorithms. First, through the application of three-dimensional fluorescence spectroscopy technology, we determined that the optimal excitation wavelength for difenoconazole is 420 nm. Next, we constructed qualification determination models using the K-nearest neighbors (KNN) algorithm and decision tree algorithm. We then selected the uninformative variable elimination (UVE) method and successive projections algorithm (SPA) as wavelength selection methods. The selected wavelengths were introduced into the broad learning system (BLS) for modeling the prediction of difenoconazole content and compared with traditional partial least squares regression (PLSR) and echo state network (ESN) models. The results indicate that the decision tree algorithm performed exceptionally well in the qualification determination model, achieving an accuracy of 97% in the prediction set. In the content prediction model, the UVE combined with BLS model exhibited excellent performance in predicting difenoconazole content, with a prediction set coefficient of determination (Rp2) of 0.959 and a root mean square error of prediction (RMSEP) of 1.358. This study has successfully demonstrated the feasibility of combining fluorescence spectroscopy technology with the broad learning system, providing a reference for the online monitoring system of pesticide residue content.

The detection of ethanol-water solution concentration plays an important role in industries, medical care, food and other aspects, which has attracted much attention. In this paper, a 632.8 nm laser combined with the oblique-incidence reflectivity difference (OIRD) method was used to obtain a signal linearly related to the solution concentration and containing the information of the dielectric constant of the solution. Combined with a variety of deep learning algorithms, ethanol-water solutions with a volume concentration of 0-95 % are detected. Among them, the prediction accuracy of the MLP, CNN, LSTM, CNN + BiLSTM + Attention models were 93.65 %, 96.54 %, 97.12 %, 99.23 %, respectively. The experimental results indicate that the OIRD method can achieve rapid, non-destructive, accurate and reliable detection of ethanol-water solutions.

Edible oil plays an important role in people's diet. Due to economic and nutritional benefits, it is necessary to analyze the components in edible oils, which can ensure consumer rights. In this study, edible oil components were identified by oblique-incidence reflectivity difference (OIRD) which included linoleic acid, oleic acid, alpha linolenic acid. In order to better analyze OIRD detection results, eight deep learning algorithms were used to analyze OIRD detection data. Experimental results suggested that all model can identify oleic acid with high Precision, Recall and F1 score. Time Series Transformer (TST) model had the highest F1 score of linoleic acid, oleic acid and alpha linolenic acid, which demonstrated best model performance. OIRD method displayed the potential for identification of edible oil components.

The wide use of deltamethrin pesticides has become a major public health issue of global concern. Due to the colorless and odorless nature of trace deltamethrin in water, the quantitative detection of its residue faces significant technical challenges. This study innovatively combines surface-enhanced Raman spectroscopy (SERS) with enhanced deep neural networks and proposes a highly sensitive and accurate quantitative analysis method for deltamethrin. The traditional CNN model was structurally strengthened by introducing the gated recurrent unit (GRU) and the Attention mechanism, and a CNN-GRU-Attention enhanced hybrid neural network was constructed. Experimental comparisons show that this enhanced model is significantly superior to the traditional PLSR, SVM and basic CNN methods in terms of feature extraction ability and nonlinear relationship modeling. Its prediction results achieve excellent performance with R2 = 0.9827 and RMSE = 0.3896. Compared with the traditional detection methods, the accuracy is increased by 8 % and the error is reduced by 40 %. Research shows that the fusion strategy based on SERS and enhanced deep learning, through multi-dimensional feature attention focusing and time-dependent modeling, effectively breaks through the sensitivity bottleneck of traditional analytical methods in trace pesticide detection, providing an innovative technical path for the precise monitoring of environmental pollutants.

Background: Ensuring the authenticity and traceability of food is fundamental to reducing food fraud, safeguarding public health, and fostering consumer trust-cornerstones of global food safety. As food supply chains grow increasingly complex, artificial intelligence (AI), in conjunction with the Internet of Things (IoT) and blockchain, plays a pivotal role in enhancing detection accuracy, improving transparency, and addressing critical challenges in food traceability. Scope and approach: This paper provided a comprehensive review of AI applications in food safety, focusing on spectroscopy, mass spectrometry, imaging, and sensor-based detection. It also examined the integration of AI with IoT and blockchain, highlighting their potential in building safe, transparent, and scalable traceability frameworks. Furthermore, the study explored how this integrated framework advanced Food Industry 4.0, driving automation, real-time monitoring, and interconnected supply chains. Finally, the paper discussed current challenges and offered perspectives on advancing AI-driven systems for food authenticity detection and traceability. Key findings and conclusions: Meanwhile, the convergence of AI, IoT, and blockchain has facilitated cross-platform compatibility and scalability, optimized supply chain data collection, and strengthened the security of traceability information. The rapid advancement of the AI-IoT-blockchain framework is driving the evolution of Food Industry 4.0, fostering advancements in high-precision analysis, automation, cost-effectiveness, and quality control, thereby enhancing food safety and transparency.

Rice is the main staple food for more than half of the world's population. Consumption of long-stored rice will have adverse effects on the human body. Here, we proposed the near-infrared (NIR) hyperspectral imaging (HSI) technique to distinguish rice from different storage years. Multiplicative Scatter Correction (MSC), Standard Normalize Variate (SNV) and 1st Derivative (1st) were used for the pretreatment of HSI data. In order to reduce dimensional spectral features, Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (tSNE) were used for data visualization. Besides, spectral characteristic wavelengths were extracted by competitive adaptive reweighted sampling (CARS) and least absolute shrinkage and selection operator (Lasso). Simultaneously, the textural features of rice were analyzed by Local Binary Pattern (LBP), Gray-Level Co-occurrence Matrix (GLCM) and Tamura algorithms. In order to realize feature fusion of spectra and texture, Support Vector Machine (SVM) model was optimized using the Whale optimization algorithm (WOA) and Extreme Gradient Boosting (XGBoost) model were established based on spectral features and textural features. Compared to other models, feature fusion model showed an excellent result with an accuracy of 98.89%. Experimental results suggested that HSI technology can be served as an effective method for rice detection.

The quality of water is paramount for human health and societal advancement. The widespread usage of pesticides, such as the herbicide bentazone, poses potential threats to water quality of sources and drinking water. This study employed Surface-Enhanced Raman spetroscopy (SERS) in conjunction with chemometric methods, aiming to provide a solution for the trace detection of bentazone residues in drinking water. Utilizing nano-silver sol as an efficient SERS substrate, SERS spectra were acquired for a total of 200 bentazone solution samples, comprising 20 distinct concentrations at ambient temperature. Feature selection and model optimization were conducted based on UVE (uninformative variable elimination), ICO (interval combination optimization), CARS (competitive adaptive re-weighted sampling) and BOSS (bootstrapping soft shinkage). These approaches significantly enhanced the accuracy and stability of the quantitative model. Simultaneously, employing the B3LYP method within density functional theory (DFT) and based on the 6-31G* (d,p) basis set, the Raman spectral characteristics of an individual molecule of bentazone were simulated and theoretically calculated. Molecular vibration modes corresponding to the characteristic peaks were analyzed. Subsequently, the correctness and interpretability of algorithms' selection of spectral region were validated based on this foundation. The results indicated that the BOSS-PLS model showed superior performance, yielded Rc2 = 0.99289, Rp2 = 0.96697, RMSEC = 0.41031 and RMSEP = 0.89701. The feature selection strategy of BOSS algorithm involved 42 variables, significantly fewer than original SERS spectra. Based on the SERS method, the final limits of detection (LOD) and quantification (LOQ) for residual bentazone in drinking water were determined to be 0.016 mg/L and 0.05 mg/L, respectively. These values were significantly lower than the standards set by the relevant standards. In summary, this study presented a novel and efficient approach by integrating SERS technology with the BOSS-PLS model, offering a feasible and reliable solution for the detection of trace residues of bentazone in drinking water.

Tea grade classification is a crucial standard in assessing tea quality, allowing for a rapid evaluation of the general quality of tea. This paper proposes a method for black tea grade classification based on hyperspectral imaging technology, chemometric algorithms, and machine learning. Standard Normal Variate (SNV) was used to preprocess the hyperspectral data. Then, correlation analysis was conducted between tea grades and chemical values to identify the chemical values that directly reflect tea grades. To reduce redundant wavelengths and improve model accuracy, Successive Projections Algorithm (SPA) and Competitive Adaptive Reweighted Sampling (CARS) were employed for secondary feature wavelength selection. Support Vector Machine (SVM) was used to establish classification models for single and multiple fused chemical values. The results show that combining hyperspectral imaging technology with machine learning can achieve excellent results in tea grade classification, with an accuracy of 88%. This method provides a theoretical basis for the rapid and non-destructive online classification of black tea grades.