engleski

Carbon nanotube stress-strain behavior prediction using convolutional neural networks

Carbon nanotubes have the potential to unlock new design possibilities and technologies. However, the computational cost of molecular dynamics simulations can be a significant barrier to determining the unique properties of various carbon nanotube structures. To obtain the most reliable and comprehensive evaluation of the mechanical properties, we predicted stress-strain curves for all single-walled carbon nanotube configurations with diameters of up to 4 nm using artificial neural networks. For the dataset obtained by molecular dynamics, two models are investigated - a multilayer perceptron and a convolutional neural network - and their performance compared. The results show that the convolutional neural network model converges faster, with a lower error. The impact of various input feature combinations on prediction quality is investigated. Interestingly, the diameter alone is a key feature for accurate prediction of thermal fluctuations. Additionally, the effect of the dataset size on prediction quality is analysed. It is shown that half of the total dataset is sufficiently representative of the problem, having a mean squared error below 1e-4 on the test set.

Single-walled carbon nanotubes, machine learning, deep learning, molecular dynamics

nije evidentirano