Physics-based Regularization of Neural Networks for Aerodynamic Flow Prediction

Abstract

Aerodynamic data plays a central role in the process of aircraft design, optimization and certification. For these processes a vast amount of data is required for various flight conditions throughout the flight envelope. Currently this data is commonly produced using Computational Fluid Dynamics (CFD). However, such simulations based on the Reynolds-averaged Navier-Stokes equations are computationally expensive and become prohibitive for tasks such as load analysis and shape optimization. During the last decades, this has motivated research focusing on the use of data-driven models with lower evaluation times than the full-order model to replace high-fidelity CFD simulations. More recently, deep learning approaches have gathered significant interest in the aerodynamic community. For the task of predicting surface pressure coefficient distributions, one of the proposed models consists of a multilayer perceptron that for each node in the mesh outputs a prediction of the local coefficient based on the node coordinates and the global operational conditions. If required, known integration formulas are used to compute integral quantities, such as the lift and pitching moment coefficients, based on the previously obtained distribution. In this paper we The method is tested for the NASA Common Research Model transport aircraft with an underlying mesh consisting of around 500,000 surface points. Results show that, when using the mentioned approach for the fine-tuning of a trained multilayer perceptron, physical knowledge can be explicitly revealed to the deep learning model but only limited improvements are achieved in the predictions of the lift and pitching moment coefficients.