Co-authors​ :

 Tim SCHMIDT (GERMANY), Dinesh NATARAJAN , Stefano CASSOLA , Miro DUHOVIC , Marlon NUSKE (GERMANY), David MAY (GERMANY) 

In the process simulation liquid composite molding, the flow of a resin through the fiber structure is governed by flow phenomena at different spatial scales spanning from microns (microscale) to meters (macroscale). A crucial step for simulation is the estimation of permeability of the fiber structure at the microscale. The permeability is a material parameter, which quantifies the conductance of a porous media for fluid flow. The microscale permeabilities are applied as homogenized properties at higher spatial scales (mesoscale, macroscale) of the porous medium. Current methods typically 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 further computational effort is necessary for flow simulations 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 are of great interest.
In this work, machine learning models were developed that can predict microscale permeabilities with a similar accuracy compared to numerical solvers, while speeding up computations at inference time. Training data for these data-driven models was generated using GeoDict [1] by modelling more than 6500 statistical representative volume elements of different microscale fiber structures and numerically computing their permeability (available on Zenodo: https://doi.org/10.5281/zenodo.10047095). The modeling data was used to develop feature-based and geometry-based surrogate models and for the comparison of the permeability predictions. In the geometry-based models, the microstructural geometry is fed into the Neural Network (NN) 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 given as input data. In addition, it is also investigated how well the models generalize and are able to predict the permeability of previously unseen fiber geometries. In the next step, the NNs are integrated into the permeability determination at the mesoscopic level. For this purpose, the mesoscopic structure is divided into suitable areas and the microscale permeability is predicted with the NNs according to their local microstructure.