Session: 36-03 Neural-Network based approaches (2)

Paper Number: 126830

126830 - Non-Parametric Surrogate Model (NPSM) for High-Dimensional Aerodynamics Optimization 

A NPSM for high-dimensional aerodynamics optimization is introduced in this paper. It directly builds a mapping relationship between surface meshes of objects and the two-dimensional (2D) distributions of flow variables. NPSM utilizes Graph Neural Networks (GNNs) to extract performance-sensitive geometric features from surface meshes according to the feedback of prediction error. With these geometric information, NPSM can predict 2D distribution of flow variables with Convolutional Neural Networks (CNNs) rather than only several performance metrics. This process bypasses manual parameterization, which reduces the uncertainties in surrogate models for high-dimensional geometries, such as fuselage, wings and blades in turbomachinery. The mapping relationship established by NPSM enables the Non-Parametric Sensitivity Analysis (NPSA) via Automatic Differentiation (AD), which can derive sensitivity of each mesh vertex rather than sensitivity of geometric parameters. This sensitivity distribution on geometry makes NPSM explainable, and provides more information for redistribution of geometry control points. To safeguard the robustness of optimization, a design classifier is designed to utilize the latent space built by NPSM to identify predictable designs. In this paper, two cases are used to demonstrate NPSM. The symmetric bumps case is used to demonstrate the methodology and functionalities of NPSM. The capability of processing high-dimensional geometries is demonstrated with the low-pressure (LP) steam turbine rotor.

Presenting Author: Jiajun Cao Whittle Lab

Presenting Author Biography: PhD in Whittle Lab

Authors:

Jiajun Cao Whittle Lab
Liping Xu Whittle Lab