A computational framework for nanotrusses: input convex neural networks approach
Abstract
The present research aims to provide a practical numerical tool for the mechanical analysis of nanoscale trusses with similar accuracy to molecular dynamics (MD). As a first step, MD simulations of uniaxial tensile and compression tests of all possible chiralities of single-walled carbon nanotubes up to 4 nm in diameter were performed using the AIREBO potential. The results represent a dataset consisting of stress/strain curves that were then used to develop a neural network that serves as a surrogate for a constitutive model for all nanotubes considered. The cornerstone of the new framework is a partially input convex integrable neural network. It turns out that convexity enables favorable convergence properties required for implementation in the classical nonlinear truss finite element available in Abaqus. This completes a molecular dynamics-machine learning-finite element framework suitable for the static analysis of large, nanoscale, truss-like structures. The performance is verified through a comprehensive set of examples that demonstrate ease of use, accuracy, and robustness.
Cite
@article{arxiv.2311.16715,
title = {A computational framework for nanotrusses: input convex neural networks approach},
author = {Marko Čanađija and Valentina Košmerl and Martin Zlatić and Domagoj Vrtovšnik and Neven Munjas},
journal= {arXiv preprint arXiv:2311.16715},
year = {2023}
}