Application of Ensemble Machine Learning Algorithms for Modeling the Thermomechanical Properties of Nano-Filled Epoxy Composites

The article is devoted to a comprehensive approach to predicting the heat resistance of plasticized epoxy nanocomposites. The aim of the study was to apply ensemble machine learning algorithms to identify key mechanical factors in predicting the heat resistance of epoxy nanocomposites and to develop a reliable method for their selection for further modeling. To achieve this goal, a two-stage hybrid methodology was applied. At the first stage, ensemble machine learning algorithms, in particular the fregression method, were used to calculate feature importance, assess their information value, and determine the consistency score between different models. In the second stage, classical statistical analysis, including Pearson's correlation and one-dimensional F-test, was performed to evaluate linear relationships. The results showed high consistency between the methods. It was found that Elastic modulus and Adhesive strength have consistently high significance (Information Value  0.15 and 0.12, respectively; Consistency Score 82% and 81%, respectively) and are key determinants of heat resistance. Conversely, Destruction stresses and Residual Stresses are completely uninformative (Information Value  0) and are recommended for exclusion from the models. Filler concentration showed minimal influence in the studied range. Thus, the effectiveness of a hybrid approach combining classical statistics and machine learning for objective and reasonable feature selection in predicting the properties of composite materials has been proven. The proposed algorithm allows the creation of simplified, interpretable, and accurate prediction models for engineering design.