Explainable AI unlocks temperature-driven oscillatory viscoelastic transitions in sesame protein isolate during integrated heating-cooling cycles

Abstract

The temperature-dependent viscoelastic behavior of sesame protein isolate (SePI) gels was investigated across integrated heating-cooling cycles (25-95 degrees C) under oscillatory rheometry (10 % strain, 0.1 Hz). Experiments were performed across a range of treatment conditions, including pressure levels of 0, 50, and 100 MPa and ionic concentrations (IC) of 0-200 mM. Empirical results showed that storage modulus (G ') consistently exceeded loss modulus (G ''), particularly during cooling, indicating elastic-dominant gelation. Application of pressure and ionic concentration (IC) treatments enhanced viscoelastic recovery, yet condition-specific nonlinear trends in G ' and G '' responses-particularly across temperature cycles-and associated hysteresis effects remained difficult to isolate from aggregated empirical trends alone. To address these limitations, stacking ensemble mimicry models were developed and explainable AI (XAI) methods, including SHAP values, partial dependence plots (PDPs), and variance-based sensitivity indices (VBSIs), were employed. The XGBMeta-Stacker and LGBMMeta-Stacker models predicted G ' and G '' with high accuracy, achieving R2 values above 0.94 for both training and testing sets. Despite variability and outliers in the temperature sweep dataset, both ensemble models showed strong predictive alignment with actual values, highlighting the robustness of the stacking strategy in complex rheological modeling. XAI analyses uncovered temperature-driven oscillatory viscoelastic transitions-repeated patterns unlikely to be captured when heating and cooling cycles are examined separately, particularly between 25 and 75 degrees C-highlighting the necessity of integrated cycle analysis to reveal such behavior and enabling quantitative ranking of temperature, pressure, and IC influences across the domain. Temperature emerged as the dominant driver of G ' and G '' transitions, while pressure exerted stronger effects on viscous behavior under high-intensity conditions. Integrated interpretation of SHAP, PDP, and VBSI analyses revealed condition-dependent feature dynamics and interaction effects, offering mechanistic insights inaccessible through traditional methods alone.