An image-based deep learning framework for flow field prediction in arbitrary-sized fibrous microstructures

Abstract

Numerical simulations are commonly used to predict resin flow in fibrous reinforcements but exhibit a trade-off between accuracy and computational cost. As an alternative, machine learning (ML) based models pose as a potential tool to accelerate or replace such costly simulations. This work proposes an open-source image-based deep learning framework to estimate the permeability of unidirectional microstructures in arbitrarily sized domains. This presents a scalable step towards estimating the permeability of large meso-domains. First, we present two robust and accurate surrogate models capable of predicting microstructure velocity and pressure fields with varying physical dimensions, fiber diameter, and volume fraction. These predictions achieve 5% error on the training set and 8% error on unseen microstructures. Secondly, based on those predicted flow fields, we infer the permeability of the microstructures with respectively 4% and 6% deviation for the training and validation sets. Third, opposed to previous works limited to microstructures with a fixed aspect ratio, we propose a so-called sliding window procedure, based on physics-based principles to predict the resin velocity and pressure field in microstructures with different aspect ratios. The method is validated against high-fidelity numerical simulations, and its predictive performance and computational efficiency are confirmed with μ-CT scans of real microstructures. Finally, the presented code and surrogate model are open-sourced to promote further exploration by the scientific community.