Modelling Relationships Between Extrusion Conditions and Quality Attributes of Expanded Snacks

Expanded snack extrusion is governed by tightly coupled interactions among raw material composition, moisture, barrel temperature, screw speed, feed rate, screw configuration, die geometry, and energy input. These variables affect not only final responses such as expansion ratio, bulk density, hardness, crispness, and water absorption or solubility indices, but also intermediate state variables including specific mechanical energy (SME), melt temperature, die pressure, melt viscosity, and bubble growth dynamics. As a result, modelling has become essential for product design, process optimisation, and scale-up. This review critically evaluates the major classes of models used to describe process–structure–quality relationships in the extrusion of expanded snacks. The literature shows that empirical regression and response surface methodology (RSM) remain the most widely applied tools because they are experimentally efficient and easy to interpret. However, mixture-process designs are more appropriate when formulation and operating variables are changed simultaneously, while phenomenological and mechanistic approaches provide better physical insight into expansion and structure development. More recently, machine-learning and interpretable artificial intelligence approaches have demonstrated strong predictive capability when large, well-curated datasets are available. Across model families, a consistent theme is that operating variables act on final product quality through intermediate process state variables rather than independently. On that basis, this review proposes a practical hybrid framework for expanded snack extrusion: a mixture-process quadratic model augmented with SME, die pressure, melt temperature and shear-related state variables, and structured in three levels linking (i) controllable inputs to state variables, (ii) state variables to measurable quality attributes, and (iii) quality attributes to a gold-standard product target or sensory-control criterion. Such a model offers a realistic balance between predictive performance, physical interpretability, experimental burden, and industrial usefulness, while also providing a clear pathway toward future digital twin and machine-learning-enabled optimisation.