Data-driven Permeability Prediction of 3D Fibrous Microstructures

Abstract

For the manufacturing process simulation of fiber-reinforced polymer composites, f low simulations have to be performed at multiple spatial scales which govern the flow through the fiber structures. Repetitive multiscale flow simulations are computationally expensive and time-consuming. In order to speed up the multiscale simulation workflow, fast machine learning surrogate models or emulators could be used to replace one or more of the flow simulations. In this work, feature-based emulators and geometry-based emulators are developed using neural networks for predicting the permeability of 3D fibrous microstructures based on a reference dataset (doi:10.5281/zenodo.10047095). The best model achieved a mean relative error of 8.33% on the test set with a significantly faster inference time compared to a conventional simulator.

Full Paper