MODELING OF ADHESIVE COMPOUNDS OF POLYMERS IN CONTACT WITH METALLS

In modern materials science, polymer-based synthetic materials are increasingly being used, including using composite materials. These materials are widely used as adhesive compounds in contact with metals. Ensuring the reliability of modern materials in harsh operating conditions determines the use of multifunctional polymer materials with a set of properties and capable of performing several important functions. At the same time, polymer materials can provide high physical and mechanical properties, adhesion, protection from aggressive media, high temperatures, vibration, atmospheric action, and possess special technological and operational properties. The article develops models for predicting the surface energy characteristics of polymers in adhesive compounds with metals. The processes of creating adhesive compounds based on polymer materials are directly related to the possibility of predicting and regulating interphase and intermolecular acid-base interactions between the components of adhesive compositions. In the course of the work, regression and classification models were built using ridge regression, the nearest neighbor method, the support vector method, artificial neural networks and the random forest algorithm, where the predicted values were the acidic and basic parameters of free surface energy and the label of the polymer belonging to the acid/basic class. It is shown that the obtained models are able to predict acidic and basic properties of polymers equally well. Linear representations of SMILES were chosen as representations of polymer molecules. Data on the surface energy characteristics of polymers were collected manually based on publications in the period from 1969 to 2022. Morgan's structural keys and molecular fingerprints were chosen as descriptors. The structure-property dependency models were built using ridge regression methods, the nearest neighbor method (kNN), the support vector machine (SVM) method and artificial neural networks using the example of a multilayer perceptron (MLP). We have shown that the k-nearest neighbor method has the best predictive ability. The statistical characteristics of the constructed kNN model for our sample have the following values: ACC = 0.89, SPC = 0.92, SEN = 0.86, BA = 0.89, Accuracy = 0.92, F1-measure=0.86.