Prediction of mechanical properties of cross-linked polymer interface by graph convolution network

Machine learning models have made significant advances in the establishment of structure-property relationships. However, it is still a challenge to predict the mechanical properties of the adhesive interface due to the complexity and randomness of the polymer topologies. In this paper, we employed a graph convolutional network (GCN) model to predict the mechanical properties of a specific cross-linked polymer interfacial system, including yield strength (σ_y), ultimate strength (σ_u), failure strain (ε_u), and fracture toughness (Γ) utilizing molecular dynamics simulations. The results showed that the adopted GCN model can predict the mechanical properties with over 88% accuracy. Furthermore, the prediction performances for ε_u and σ_u are better than those for Γ and σ_y, with R² ∼ 0.73 for ε_u, R² ∼ 0.64 for σ_u, R² ∼ 0.51 for Γ, and R²∼ 0.43 for σ_y. It is worth noting that the GCN model with the sum aggregator slightly outperforms that with the mean aggregator, and that models with linear regression and fully connected neural network regression provide similar predictions. The influence of input node features on prediction performance was also investigated. It was observed that the node closeness centrality is an important graph parameter in prediction. Specifically, node closeness centrality presents a more significant influence on the global mechanical properties of the adhesive interface, such as ε_u, σ_u, and Γ. Additionally, sensitivity analysis demonstrated that appropriate hyperparameters can improve computational efficiency without losing accuracy on a restricted set of data. This paper demonstrated the capacity of the GCN model to predict the mechanical properties of the adhesive interface with diverse topologies and provided a possible pathway for improving the mechanical properties of the adhesive interface by tailoring polymer structures in the future.