Fast Flow Field Estimation for Various Applications with A Universally Applicable Machine Learning Concept
Michael Leer, Andreas Kempf
Abstract
Abstract This paper presents an approach for the prediction of incompressible laminar steady flow fields over various geometry types. In conventional approaches of computational fluid dynamics (CFD), flow fields are obtained by solving model equations on computational grids, which is in general computationally expensive. Based on the ability of neural networks to intuitively identify and approximate nonlinear physical relationships, the proposed method makes it possible to eliminate the explicit implementation of model equations such as the Navier–Stokes equations. Moreover, it operates without iteration or spatial discretization of the flow problem. The method is based on the combination of a minimalistic multilayer perceptron (MLP) architecture and a radial-logarithmic filter mask (RLF). The RLF acts as a preprocessing step and its purpose is the spatial encoding of the flow guiding geometry into a compressed form, that can be effectively interpreted by the MLP. The concept is applied on internal flows as well as on external flows (e.g. airfoils and car shapes). In the first step, datasets of flow fields are generated using a CFD-code. Subsequently the neural networks are trained on defined portions of these datasets. Finally, the trained neural networks are applied on the remaining unknown geometries and the prediction accuracy is evaluated. Dataset generation, neural network implementation and evaluation are carried out in MATLAB. To ensure reproducibility of the results presented here, the trained neural networks and sample applications are made available for free download and testing.