Computational Fluid Dynamics Analysis Surrogates Based on Polynomial Regression and Convolutional Neural Network Machine Learning

Abstract

Computational fluid dynamics (CFD) is the standard method for simulating fluid flow. The main drawback to CFD is the high computational cost to perform a single analysis. The objective of this work is to develop fast methods for computing an estimate of the system model, given pre-computed component solutions and system specifications. This paper explores surrogates for CFD analysis for use in parametric exploration of the design space on the corpus of components and seed designs. The surrogates considered are polynomial regression and convolutional neural network (CNN) machine learning algorithms. The polynomial regression surrogate uses the dimensions of the geometry under consideration to predict the drag force, lift force, and moment. The polynomial regression model performed well but did not generalize to unseen geometries. The CNN surrogate uses pre-computed pressure contours from each component in the geometry under analysis to predict the surface pressure on the aggregation of components. This approach allows surface pressure predictions for parametric exploration in a matter of minutes. The CNN surrogate predicted surface pressure values on geometries seen in the training set with 93.96% of the predictions having less than 25% error and unseen geometries with 76.91% of the predictions having less than 25% error. Additionally, the CNN surrogate demonstrated a 3.5 times decrease in execution time compared to traditional CFD methods.