Comparative Evaluation of Machine Learning Algorithms for Predicting the Mechanical Properties of Polymer Composites

This paper presents the results of an investigation into the feasibility of applying ensemble machine learning algorithms to predict the mechanical properties of nanofilled epoxy composites, aimed at enhancing the reliability of water transport facilities. The study is positioned as a validated proof-of-concept demonstrating the applicability of ensemble ML models for property prediction tasks in epoxy composites, rather than as the formulation of a universal predictive framework. The primary objective of this study is to develop an efficient tool for predicting the mechanical performance of nanofilled epoxy composites based on ensemble machine learning paradigms to improve the operational reliability of marine and watercraft components. Utilizing experimental data on the properties of composites formulated from DER-331 epoxy resin with varying loadings of multi-walled carbon nanotubes (0–0.125 wt.%), seven ensemble models were developed employing Decision Trees, Random Forest, Gradient Boosting, AdaBoost, Bagging, Stacking, and Voting algorithms. It was established that the AdaBoost algorithm provides the highest predictive accuracy (MAPE = 8.02%). Linear statistical tests (Pearson correlation and one-way F-test) identified a statistically significant influence only for adhesive bond strength (solidity), whereas mutual information analysis unveiled pronounced non-linear dependencies for filler loading (filler), residual stresses (residual stresses), and heat resistance (heat resistance). Adhesive strength (mean importance of 0.47; Consistency Score of 81.37%) and residual stresses (Consistency Score of 85.17%) were identified as the most consistent and critical factors across the evaluated models. In contrast, filler loading exhibited the lowest inter-model consensus (Consistency Score = −20.59%), directly reflecting the complex, non-linear nature of its reinforcing mechanism. It is proven that the implementation of ensemble machine learning algorithms coupled with subsequent feature importance analysis offers an effective framework for predicting the mechanical properties of epoxy composites, enabling a substantial reduction in the required volume of routine experimental testing. The obtained results establish a solid foundation for developing durability prediction methodologies for protective coatings utilized in water transport and can be directly applied to optimize composite formulations with tailored, predictable operational properties.