Prediction of the mechanical properties of graphene/copper nanocomposites using machine learning and molecular dynamics simulations

Doping graphene into copper monomers significantly enhances their mechanical properties, thereby broadening the application scope of graphene/copper nanocomposites. Molecular dynamics (MD) simulation serve as a powerful tool for investigating the mechanical behavior of these nanocomposites. This study systematically explores the influence of four critical factors-external temperature, graphene vacancy defects, graphene chirality, and insertion angle-on the performance of graphene/copper nanocomposites. However, the simultaneous analysis of these factors through MD simulations substantially escalates computational demands. To address the computational bottleneck of MD simulations in analyzing multifactorial interactions, we integrate LSTM networks and back propagation (BP) neural networks for dual-task prediction: (1) LSTM captures the complete tensile stress-strain behavior (300 time steps per case) by learning sequential MD data, and (2) BP networks predict Young's modulus and yield strength from critical parameters (temperature, chirality, vacancy defects). Results demonstrate that the LSTM model achieves<i>R</i>²= 0.96 for Young's modulus and<i>R</i>²= 0.94 for yield strength prediction, while the BP neural network further improves accuracy to<i>R</i>²= 0.97 for both properties. Notably, the LSTM model predicts the entire tensile process in 2.4 s per curve, reducing computational time by three orders of magnitude compared to MD simulations (typically requiring hours). Furthermore, LSTM effectively helps elucidate the whole tensile process of the composites, which enhances the ability to predict material properties.