Machine Learning for Reduced-Order Modeling of Composites Processing

Abstract

Machine learning concepts have made their way into industry practice and can offer quick and accurate solutions for various design and in-service applications including process simulation of composites. High-fidelity finite element simulation tools are currently used for thermo-chemical analysis of parts during processing. This research investigates reduced-order modeling of complex composite parts using machine learning methods to speed-up the simulation. At first, process simulations of representative 2-D stringer geometries were conducted using finite element analysis to extract leading and lagging responses. Equivalency of exothermic responses of 2-D simulations to 1-D finite element analysis was established using an in-house developed machine learning framework at the University of Washington (CompML). It is demonstrated that for a given 2-D geometry, a trained NN can identify an equivalent 1-D thermal stack to yield similar exothermic responses. The results and methods developed in this study can significantly reduce the computational cost of finite element simulation by successfully establishing reduced-order models of complex geometries.