Numerical data generation for building machine learning models for permeability estimation of fibrous structures

Abstract The challenge of developing effective multiscale flow simulation methods is to take into account all relevant physical effects from fiber to component level, but remain manageable in terms of computational effort. A challenge that can potentially be met by incorporating data‐driven machine learning methods in the multiscale approaches, for example, by substituting flow simulations at the microscale where the simulation domains have high geometric similarity. 4284 models of fibrous structures with several hundred individual fibers and varying structural parameters were generated for permeability estimation. Various neural network architectures were trained with this data and different input data strategies were compared concerning accuracy in the estimation of the direction‐dependent permeability tensor. The best case showed a mean relative error of 11.35%, with the inference time being orders of magnitude lower than the flow simulation computing time. Highlights Statistical volume elements essential for accurate permeability estimation. Large data set FiberStructurePerm generated and available on Zenodo. Permeability can be predicted well with data driven emulators. Time required to predict the permeability is reduced by over 90%.