Yuchen Ban, Yuqiao Ren, Wenrui Dong, Da-Wen Sun, Brijesh K. Tiwari
Background Rapid, accurate, and intelligent quantification of food components is essential for food quality control, nutritional evaluation, and safety monitoring. Conventional analytical methods are often destructive, time-consuming, and costly. Nondestructive spectroscopic techniques, such as terahertz time-domain spectroscopy, provide efficient alternatives; however, their industrial deployment is hindered by high-dimensional spectral data, light-scattering effects, and the complexity of food matrices. Scope and approach This paper presents a scoping and critical review of recent advances in applying deep neural network frameworks (DNNFs) to improve quantitative analysis with nondestructive spectroscopy in food systems, with emphasis on studies published since 2020. It is also explained how DNNFs can be used to support preprocessing, representation learning and regression, and the transition from hybrid pipelines to end-to-end modelling is discussed, as well as the challenges in validation, interpretability and industrial deployment. Key findings and conclusions Advanced DNNFs can show advantages over traditional chemometric methods in specific scenarios by automatically learning spectral characterization and capturing complex nonlinear relationships, especially in cases with prominent nonlinear structures, significant spatiotemporal dependence, or multi-component mixed systems. These capabilities significantly improve prediction accuracy, robustness, and model generalisation across diverse food matrices. End-to-end DNNFs-based strategies further reduce dependence on manual preprocessing and provide a promising pathway for real-time, intelligent monitoring of food quality in complex industrial environments. Continued development of scalable, interpretable, and robust DNNFs is expected to accelerate the practical implementation of nondestructive spectroscopic quantification in the food industry.