Predictive Modeling of Food Extrusion Using Hemp Residues: A Machine Learning Approach for Sustainable Ruminant Nutrition

Predictive modeling of extrusion processes through machine learning (ML) offers significant improvements over classical response surface methodology (RSM) when addressing nonlinear and multivariable systems. This study evaluated hemp residues (Cannabis sativa) as a non-conventional ingredient in ruminant diets and compared the performance of polynomial regression models against several ML algorithms, including artificial neural networks (ANNs), random forest (RF), K-Nearest neighbors (KNN), and XGBoost. Three experimental datasets from previous extrusion studies were concatenated with new laboratory experiments, creating a unified database in excel. Input variables included extrusion parameters (temperature, screw speed, and moisture) and formulation components, while output variables comprised expansion index, BD, penetration force, water absorption index and water solubility index. Data preprocessing involved robust z-score detection of outliers (MAD criterion) with intra-group winsorization, followed by normalization to a [−1, +1] range. Hyperparameter optimization of ANN models was performed with Optuna, and all algorithms were evaluated through 5-fold cross-validation and independent external validation sets. Results demonstrated that ML models consistently outperformed quadratic regression, with ANNs achieving R2 > 0.80 for BD and water solubility index, and RF excelling in predicting solubility. These findings establish machine learning as a robust predictive framework for extrusion processes and highlight hemp residues as a sustainable feed ingredient with potential to improve ruminant nutrition and reduce environmental impacts.