Fiber-reinforced polymer composites are a class of composite materials in which fibers are dispersed in a continuous polymer matrix. In the process simulation of fiber-reinforced composite manufacturing, the flow of a liquid polymer through the fiber structure is governed by flow phenomena at different spatial scales spanning from micrometers (microscale) to meters (macroscale). A crucial step in this simulation process is the estimation of permeabilities of the fiber geometry on the microscale. These microscale permeabilities are then used to homogenize the porous medium at higher spatial scales (mesoscale, macroscale). 

Current conventional methods compute the permeability of a microstructure by solving the Stokes equation that governs the fluid flow through the microstructure. However, repetitive modeling of microscale geometries and flow simulations require high computational effort and time. Since higher efforts will be necessary for the flow simulation at higher spatial scales, a fast surrogate model for the microscale permeability prediction is desirable. Here, modern machine learning and deep learning methods which offer fast inference times have become of great interest.  

In this work, data-driven emulators were developed using machine learning that can predict microscale permeabilities with a similar accuracy as a commercial solver GeoDict, while speeding up the computations at inference time. Training data samples for these data-driven models were generated using GeoDict unit cell models of synthetically generated microscale fiber structures and computing their permeability in various directions. These data samples were used to compare two categories of emulators – feature-based and geometry-based models. In the geometry-based models, the microstructure geometry is fed as input data, whereas in the feature-based models, only the effective parameters characterizing the microstructure geometry such as fiber volume content, fiber diameter and fiber orientation are fed as input data. It is also investigated how well the models generalize in predicting permeabilities of previously unseen fiber geometries.