A computational framework for nanotrusses: enter convex neural networks method
Authors: Marko Čanađija, Valentina Košmerl, Martin Zlatić, Domagoj Vrtovšnik, Neven Munjas
Summary: The current analysis goals to supply a sensible numerical device for the mechanical evaluation of nanoscale trusses with related accuracy to molecular dynamics (MD). As a primary step, MD simulations of uniaxial tensile and compression checks of all potential chiralities of single-walled carbon nanotubes as much as 4 nm in diameter had been carried out utilizing the AIREBO potential. The outcomes signify a dataset consisting of stress/pressure curves that had been then used to develop a neural community that serves as a surrogate for a constitutive mannequin for all nanotubes thought-about. The cornerstone of the brand new framework is {a partially} enter convex integrable neural community. It seems that convexity permits favorable convergence properties required for implementation within the classical nonlinear truss finite ingredient obtainable in Abaqus. This completes a molecular dynamics-machine learning-finite ingredient framework appropriate for the static evaluation of enormous, nanoscale, truss-like buildings. The efficiency is verified by means of a complete set of examples that exhibit ease of use, accuracy, and robustness.
