Title: Reduced-order model for multiphysics simulations of CNT/Polymer Composites via principal component regression and artificial neural networks

Kavan Shah, Krishna Kiran Talamadupula, Pinar Acar and Gary D. Seidel
Department of Aerospace and Ocean Engineering
Virginia Tech
Blacksburg, VA 24060

Computational Materials Science Volume 244, September 2024, 113200

Abstract
In this work, the stochastic microstructure of simulated CNT-polymer composite statistical volume elements (SVEs) is quantified using two-point correlation functions. The two-point correlation functions are represented in a lower-dimensional subspace obtained via Principal Component Analysis (PCA) to make the correlation functions more tractable for use in Machine Learning algorithms for property prediction. Linear regression and Artificial Neural Network (ANN) models are used to establish structure–property linkages that predict the effective stiffness, electrical conductivity, and piezoresistivity coefficients from the principal scores of the correlation functions. It is found that the linear regression and ANN models predict the effective stiffness accurately, but fail to predict the electrical and electro-mechanical properties with sufficient accuracy. It is inferred that the lower-dimensional principal space representation of the two-point correlation functions captures the CNT volume fraction and CNT orientation distribution, both of which control the effective stiffness of the SVEs. However, this representation fails to adequately capture important factors which control the electrical conductivity and piezoresistivity of the SVEs, which include the formation of conductive paths between CNTs and the sensitivity of these paths to perturbations in the positions of CNTs due to applied strain.

Key words: Carbon nanotube Statistical characterization Piezoresistivity Two-point correlation functions Principal component analysis Neural networks Multiphysics surrogate model