Predicting stress-strain behavior of carbon nanotubes using neural networks

Artificial neural networks are employed to predict stress-strain curves for all single-walled carbon nanotube configurations with diameters up to 4 nm. Three model architectures are investigated for the molecular dynamics-derived dataset: a multilayer perceptron, a one-dimensional convolutional neural network, and a residual neural network. The performance of the three models is compared, and they are found to closely match an atomistic-physics-based paradigm while being orders of magnitude faster. The effect of the dataset size on the prediction quality is analyzed. It is shown that 30% of the entire carbon nanotube configuration dataset is representative of the problem. Remarkably, all models demonstrate high accuracy, capturing even the smallest variations due to thermal fluctuations, and can provide averaged stress-strain curves without thermal fluctuations. Additionally, a sensitivity analysis was performed to investigate how the various input feature combinations affect the quality of elimination or prediction of thermal fluctuations. The results are determined by different combinations of input features, with current diameter in combination with temperature identified as the most important parameters affecting the inclusion or exclusion of thermal fluctuations.

Automated summary

Artificial neural networks were used to predict stress-strain curves for carbon nanotubes of different configurations. The study compared three model architectures and found that they can accurately and quickly predict the curves, capturing even the smallest variations caused by thermal fluctuations. A sensitivity analysis revealed that the diameter and temperature are the most important parameters affecting the exclusion or prediction of thermal fluctuations.