A Weibull dose-response descriptor and nested regression framework for predicting radiation-induced gelation and downstream properties in electron-beam irradiated LLDPE/EBC blends

This study develops a two-stage physics-guided machine-learning workflow for electron-beam irradiated LLDPE/EBC blends. In the first stage, gel content was modelled from blend composition and irradiation dose, while leave-one-polymer-system-out (LOPSO) and within-system leave-one-dose-out (LODO) validation addressed distinct generalisation tasks. In a common-fold LODO comparison, empirical Weibull, Charlesby-Pinner (CP), Charlesby-Pinner-Rosiak (CPR), and training-mean baseline models were evaluated on the same 36 held-out gel predictions. Weibull had the lowest pooled OOF error, with MAE 4.386 and RMSE 6.961 percentage points, versus 7.911/11.123 for CP and 8.882/12.927 for CPR. CP and CPR were retained as theory-based comparators; CPR was boundary-dependent and its fitted parameters were not interpreted as radiation yields, molecular descriptors, or physical virtual doses. For the primary fully unseen-blend scenario, downstream models used LLDPE composition and dose, with an outer-training-only gel surrogate refitted within the outer and inner grouped folds. Paired outer-LOPSO MAE/RMSE/bias were 1.467/1.694/0.191 and 1.417/1.578/0.190 Shore D for process-only and physics-informed hardness models, respectively; the corresponding heat-shrinkage values were 3.114/3.760/0.981 and 2.869/3.401/1.081 percentage points. The pooled RMSE reductions were not uniform across held systems and their six-system bootstrap intervals quantify predictive stability, not experimental measurement uncertainty. The findings support a fixed-grade proof of concept and do not remove the need for condition-matched molecular-weight, crystallinity, morphology, interfacial, and phase-size measurements for transferable blend-design claims.