Data‐Driven Design of Epoxy Composites: Heat Resistance Prediction Using Machine Learning Algorithms

The paper presents the results of studies of the mechanical properties of modified and dispersion‐filled epoxy composite materials: destructive stresses and modulus of elasticity in bending, impact toughness, adhesive strength, residual stresses and heat resistance. At the initial stage, we applied the analysis of a set of properties only for training ensemble machine learning (ML) algorithms. Based on experimental data, we then considered the problem of predicting the values of the variable heat resistance depending on the values of the mechanical properties of composites. To solve the prediction problem, we used various ML algorithms: AdaBoost, bagging, decision trees, gradient boost, random forest, stacking, voting, ridge regression, support vector machine (SVM) and K‐nearest neighbour ( k ‐NN), which establish the dependence of the values of the variable heat resistance on the values of the above variables. To ensure the reliability of the results and minimise dependence on the specifics of individual algorithms, a comprehensive approach was used, combining ML and mathematical statistics methods. The results of solving two important tasks have been obtained: the first task is to predict the values of the heat resistance variable based on the values of the mechanical properties of materials, and the second task is to determine the importance of the influence of the values of all variables on the value of the predicted heat resistance variable. Based on the results of the first task, it can be stated that all models (ML algorithms) obtained in the work for prediction are highly effective since their MAPE has low values on the test sample and, therefore, can be used by experimenters to reduce the costs of various types of resources for conducting experiments. Based on the results of solving the second task using three different approaches, conclusions can be drawn about the significance of the results obtained, given that the models assess the importance of the influence of variable values in the context of their complex interaction with each other. The information obtained has potential value for experimenters in their optimal management of variable importance values and in organising experiments to create new materials with predicted values of the selected variable.