AI-guided extraction of plant-derived bioactives for edible packaging films

This is the authors' abstract. We don't add key points for this paper.

The increasing concern for plastic pollution and the rising demand for sustainable food packaging, have accelerated research on edible packaging films fortified with bioactive compounds from plants. Phytochemicals such as flavonoids, terpenoids and polyphenols confer antioxidant, antimicrobial and functional properties to improve food safety, maintain quality and extend shelf-life. However, conventional extraction methods such as Soxhlet, maceration and solvent extraction have disadvantages of high solvent and energy consumption, low selectivity and extraction efficiency and thermal degradation of bioactives. Furthermore, the nonlinear relationships among the extraction variables are complex and limit the performance of conventional statistical optimization methods. This review critically examines the integration of conventional statistical optimization tools, especially response surface methodology (RSM) and design of experiments (DoE), with machine learning (ML) and deep learning (DL) strategies for advanced green extraction processes such as microwave-assisted, ultrasound-assisted, enzyme-assisted, supercritical fluid, pulsed electric field, and pressurized liquid extraction. RSM/DoE helps in systematic experimental design and optimization of processes, whereas ML methods, including Artificial Neural Network (ANN), Support Vector Regression (SVR), Random Forest (RF), and Adaptive Neuro-Fuzzy Inference System (ANFIS), enable data-assisted prediction of nonlinear relationships. DL architectures, including Convolutional Neural Networks (CNNs), autoencoders (AE), transformers, and Physics-Informed Neural Networks (PINNs), support complex and high-dimensional datasets when sufficient training data are available; However, their significance in edible packaging research, where datasets are relatively small and heterogeneous, must be addressed in terms of model complexity, overfitting and validation robustness. It further evaluates AI-guided incorporation of plant bioactives into edible packaging through optimized encapsulation, prediction of mechanical and barrier properties, controlled release behavior, safety assessment, and sustainability evaluation. AI-assisted and hybrid models have demonstrated, in various reported studies, enhanced predictive accuracy and optimization efficacy compared with conventional response surface methodology in terms of extraction yield, solvent use, energy demand and experimental trials. However, the limitations of standardised datasets, computational complexity, regulatory issues and lack of large-scale validation restrict the applications in industry. The reviewed literature still provides limited evidence directly linking mechanical, barrier and release properties, as well as safety outcomes such as migration, to ML-based modeling and optimization. Formulation-specific experimental validation is therefore still required. Future advances in explainable AI frameworks, harmonized databases, industrial-scale studies and multidisciplinary collaboration will be crucial for the commercialization of intelligent and sustainable edible packaging systems. Further scale-up studies are needed before these approaches can be considered viable for industrial application. Their translation will depend on process reproducibility, regulatory acceptance, sufficiently large and representative datasets, appropriate computational resources, formulation-specific testing, mechanism-informed modeling and uncertainty quantification. Any proposed formulation will also require independent experimental validation, together with food-safety and shelf-life studies, before it can be accepted for food-contact applications.