Knowledge integrated, deep neural network-based prediction of stress-strain curves of polymer matrix composites for AI-assisted materials design

In order to achieve an effective energy transition, development of new materials must be accompanied with the development of new and renewable energy facilities. However, to this day, material design is costly because material development relies on the designer's intuition. Therefore, for the competitiveness of material development, AI-based material design automation must be made through the combination and composition prediction of components. As the first step in the AI-based material reverse engineering system, this study predicts the mechanical properties and behavior of polymer matrix composites (PMC). The mechanical behavior of a material can be expressed from the strain-stress curve (S-S curve), and the deformation from the elastic section to the plastic section can be judged along with mechanical properties such as tensile strength, elastic modulus, and maximum load. Therefore, this study aims to predict the mechanical behavior of the PMC by learning the minimum tensile test data and information on the components for the two-component PMC based on the deep learning methodology. Through literature/data analysis, most features that can affect mechanical properties were classified into two predictive models. The first predictive model inputs tensile test data and chemical/mechanical properties, and outputs mechanical properties behavior. And the second prediction model predicts by inputting structural information of each components. Through SMILES of each components, MACCS key was obtained and converted to use functional group information and used as a feature. As a result of comparing the performance of the two predictive models, the second model required less material information than the model that did not learn structural information, and performed better. As a result, it is a model that predicts the behavior of the plastic section beyond the existing prediction model that stayed in the elastic modulus section.