Naslov
Mechanical Properties of Carbon Nanotubes - Deep Learning Approach

Autori
Čanađija, Marko

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
NT21: International Conference on the Science and Application of Nanotubes and Low-Dimensional Materials / - Houston (TX), 2021

Skup
21st International Conference on the Science and Application of Nanotubes and Low-Dimensional Materials (NT 2021)

Mjesto i datum
Houston (TX), Sjedinjene Američke Države, 06.06.2021. - 11.06.2021

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
Carbon nanotubes ; Machine learning ; Mechanical properties

Sažetak
The research at hand presents a thorough investigation into mechanical properties of single-walled carbon nanotubes (SWCNT). The Young's modulus, ultimate tensile stress and strain at fracture were computed by the molecular dynamics (MD) from a comprehensive set of tensile tests. AIREBO potential was used, while the elongation of SWCNT was performed in the constant strain regime. Length/diameter ratio of all SWCNT was kept at L/D=5. To the best knowledge of authors, for the first time, all possible SWCNT configurations – armchair, zigzag and chiral within the diameter range 0.36-4.00 nm and chiral angles 0-30 were considered. Results show that chiral nanotubes that are structurally the closest to zigzag SWCNT exhibit the highest Young's modules, while the lowest modules were obtained in armchair SWCNT, the smallest one in particular. On the other side, armchair and structurally closest chiral SWCNTs have the highest ultimate tensile stress and strain at fracture. In general, nanotubes of smaller radius have somewhat lower ultimate tensile stress, while the fracture strain is rather uniformly distributed for a chosen chirality. Instead of the usual representation in terms of diameter as the variable, the results herein are provided in terms of chiral indices. Although the described method provided a clear illustration of SWCNT mechanical properties, the question of how to provide a sound, but approximate model of the described properties remains open. To this end, deep learning methodology is employed to provide an artificial neural network that approximate the Young’s modulus. A feedforward neural network composed of the following layers was used: 128 neurons, linear activation function ; dropout layer with the rate 0.4 ; 64/ReLU ; 64/ ReLU ; 32/ReLU ; 32/ ReLU ; 32/ ReLU ; 16/ ReLU ; 16/ ReLU ; 8/ ReLU ; 8/ReLU ; 4/ReLU ; 1 linear. Momentum optimizer with learning rate 0.1 was exploited, while the number of epochs was 30000. L2 regulation was used in hidden layers. The final value of R2 for the approximation of Young’s modulus was 0.92. This indicates a good fit to data obtained by MD and justifies the approach.

Izvorni jezik
Engleski

Znanstvena područja
Strojarstvo, Temeljne tehničke znanosti, Interdisciplinarne tehničke znanosti

POVEZANOST RADA