Correlation between starch physicochemical properties and pellet feed quality under different starch sources for mechanistic understanding and prediction

Starch is the major energy source and natural binder in pelleted feed, and its botanical origin-dependent properties strongly affect feed processing quality. This study aimed to clarify how starch physicochemical properties govern pellet forming performance and to establish machine learning (ML)-based prediction models linking starch characteristics to pellet quality. Eight starch sources (wheat, common corn, waxy corn, potato, sweet potato, manioc, pea, and mung bean) were analyzed for chemical composition, granule morphology, gelatinization, rheological, and thermal properties. Amylose content ranged from 1.05% to 41.27%, and the amylopectin/amylose ratio varied from 0.94 to 74.90. Average particle sizes were 12.92–38.71 μm, and gelatinization temperatures ranged from 69.43 to 89.58 °C. Based on significant differences in gelatinization temperature (a critical processing-related index), four representative starches (wheat, common corn, manioc, and pea) were selected for pelleting trials. Within 10–20% starch addition, pellet density and hardness were 1.32–1.86 g/cm³ and 3.18–3.96 kg, respectively. Higher starch addition significantly increased pellet density and hardness. Low gelatinization temperature facilitated densification, while high amylose favored gel network formation. Five ML algorithms (GBDT, KNN, PLS, RF, and SVR) were evaluated. GBDT performed best for density (R² = 0.93), whereas PLS performed best for hardness (R² = 0.94) and powder content rate (R² = 0.98). SHAP analysis indicated that starch addition was the dominant factor, positively contributing to density and hardness but negatively to powder content rate. This work provides predictive tools for starch source selection and formula optimization in pelleted feed.