Stochastic integrated machine learning based multiscale approach for the prediction of the thermal conductivity in carbon nanotube reinforced polymeric composites

Abstract

We present a stochastic integrated machine learning based multiscale approach for the prediction of the macroscopic thermal conductivity in carbon nanotube reinforced polymeric composites (CNT-PCs). Seven types of machine learning models are exploited, namely Multivariate Adaptive Regression Splines (MARS), Support Vector Machine (SVM), Regression Tree (RT), Bagging Tree (Bag), Random Forest (RF), Gradient Boosting Machine (GBM) and Cubist. They are used as components of stochastic modeling constructing the relationship between all uncertain inputs variables and the output of interest, the macroscopic thermal conductivity of the composite. Particle Swarm Optimization (PSO) is used for hyper-parameter tuning to find the global optimal values leading to a significant reduction in the computational cost. We also analyze the advantages and disadvantages of various methods in terms of computational expense and model complexity. We believe that the presented stochastic integrated machine learning approach accounting for uncertainties is a valuable step towards computational design of new composites for application related to thermal management.