We used CORSIKA to simulate extensive air showers initiated by primary protons and iron nuclei at an energy of 30 TeV and a zenith angle of 0∘. By analyzing secondary particles within 0-600 m from the shower axis, we extracted statistical-moment features for the electromagnetic and muonic components, and fitted the radial distributions of both components with the NKG function to obtain the age parameters and particle-number parameters (Georgios in EPJ Web Conf. 137:13001, 2017). We first constructed a stacking ensemble model (with XGBoost, Random Forest (RF), and Support Vector Machines (SVMs) as base learners, and Logistic Regression as the meta-learner) for proton-iron classification, and then introduced single classifiers (SVM, XGBoost, and Random Forest) as benchmarks to validate the reliability of the stacking framework. The area under the ROC curve (AUC) and the Q-factor were used as evaluation metrics for composition discrimination. The results show that, under this idealized proton-iron setting, the single classifiers already reach near-saturated performance (AUC close to 1). The stacking model achieves an AUC of 0.995 on an independent test set, with a maximum Q ≈1 × 10 $\mathtt{^{4}}$ in the threshold scan. The comparable performance between stacking and the single models indicates that the stacking framework is stable and reliable for this task; the lack of a significant performance gain is mainly due to the idealized physical setup and the strong class separability, which leave limited room for improvement beyond the single-model ceiling. This workflow can serve as a baseline for future composition-discrimination studies that incorporate detector response, noise/systematic uncertainties, and broader energy and zenith-angle ranges.
