A computational study for batch-level diagnostics in industrial extrusion using interpretable machine learning and counterfactual analysis

Industrial feed extrusion involves complex interactions between raw materials, processing conditions, and product outcomes. This study develops machine learning models to predict critical process and product attributes across key stages of extrusion, conditioning, extrusion, and drying, based on a large-scale, two-year dataset with high batch and recipe variability. Ensemble methods and neural networks are applied to model outcomes such as specific mechanical energy, die pressure, bulk density, hardness, and the chemical composition of final products. Model performance was evaluated using cross-validation and test data, with XGBoost achieving the highest predictive accuracy. R 2 values exceeded 0.90 for specific mechanical energy and compositional attributes and were around 0.80 for physical pellet characteristics and die pressure. Beyond prediction, model interpretability was addressed using Shapley Additive Explanations to uncover key variable interactions and support process transparency. These insights enabled simulation of counterfactual scenarios to assess how adjusting upstream parameters could influence product outcomes. The approach demonstrates how machine learning can support targeted batch optimization and reverse engineering in industrial settings, offering a scalable framework for integrating data-driven modelling into complex manufacturing workflows.