Application of machine learning method to predict debonding in graphene-reinforced polymer nanocomposite

In nanocomposites, debonding originates at the interface between reinforcing particles and the matrix. Since experimental characterization and computational modeling of particle debonding are highly resource-intensive, the development of efficient predictive strategies is essential. In this study, a machine learning framework is proposed to predict debonding evolution using data obtained from finite element simulations. A multiscale modeling approach using representative volume elements (RVEs) was implemented to analyze debonding in the interface area of polyethylene–graphene nanocomposites. Debonding at the interface between graphene nanosheets and the polymer was modeled using a cohesive-zone model defined by traction–separation curves. Cohesive-zone parameters for polyethylene-graphene were sourced from molecular dynamics simulations. RVEs were generated with a Python script and analyzed using the Abaqus explicit solver. The study examined the effects of volume fraction, aspect ratio, and loading conditions on particle debonding. The results closely matched the published literature and effectively captured the trends in debonding behavior. A dataset comprising input parameters (e.g., volume fraction, aspect ratio) and the corresponding output (debonding percentage) was fed into a 10-layer artificial neural network. The network was trained using a gradient descent algorithm with various activation functions. The trained model demonstrated reasonable generalization capabilities with minimal overfitting. Subsequently, the model was used to predict debonding percentages for unseen input data. The results accurately captured the trends in debonding behavior for varying volume fractions and aspect ratios. The observed discrepancies were attributed to the limited dataset and the absence of debonding in specific samples.