Hybrid machine learning strategies for predicting for jute/ flax/ZnO reinforced composites

This study proposes an integrated experimental and machine-learning framework for the modeling and prediction of the properties of sustainable hybrid polymer nanocomposites reinforced with natural resource jute and flax fibers and Zinc Oxide (ZnO) nanoparticles. The effects of three key processing parameters fiber orientation, fiber content, and ZnO concentration on the tensile and flexural properties of the composites were systematically investigated. Response Surface Methodology (RSM) was utilized to establish quadratic predictive models for both mechanical responses, demonstrating strong in-sample agreement with experimental observations (R² > 0.98) and providing valuable insights into the interactions among the governing variables. To further enhance predictive capability, four machine-learning algorithms, namely Decision Tree (DT), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Random Forest (RF), were developed and evaluated alongside RSM using a single, leakage-free leave-one-design-point-out nested cross-validation protocol grouped by the 15 replicated Central Composite Design points. Model performance was assessed through the coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE). Under this protocol, RSM and XGBoost gave the most accurate held-out predictions for both responses (pooled R² ≈ 0.90) and were not significantly different from one another (paired tests, p > 0.8), while Decision Tree and Random Forest showed materially lower held-out accuracy (pooled R² ≈ 0.72–0.75) and SVM was the weakest model evaluated (pooled R² ≈ 0.51–0.53). Paired tests and bootstrap confidence intervals over design points showed that Decision Tree, Random Forest and SVM were significantly less accurate than RSM ( p < 0.05). Random splitting that places replicate specimens of one design point in both training and test sets was avoided because it inflates apparent accuracy. The results revealed that fiber orientation is the most influential parameter affecting composite mechanical behavior, followed by fiber loading and ZnO nanoparticle content. The combined use of experimental design, statistical modeling, and leakage-free machine-learning validation provides a transparent framework for comparing prediction models for natural fiber-reinforced polymer nanocomposites, within the scope of the factors, dataset size, and material system examined here. Within the studied factor ranges, RSM and XGBoost are candidate prediction tools. All replicates of each design point came from a single panel, and no independent external validation was performed; the quadratic RSM model shows a significant lack of fit, and the results should not be extrapolated beyond the tested ranges or to other material systems.