Improved data-driven surrogate models by incorporating variable sensitivity for aerodynamic data modeling

doi:10.1080/19942060.2025.2564830

Improved data-driven surrogate models by incorporating variable sensitivity for aerodynamic data modeling

2025-10-09

0

OA

AI

DOI：10.1080/19942060.2025.2564830

Original

Original request for help

Save

Abstract

En 中文

Data-driven surrogate models have become increasingly important in aerospace engineering for the rapid prediction of aerodynamic characteristics. However, when modelling aerodynamic data with varying flight conditions and complex shape parameters, traditional surrogates – such as kriging and fully connected neural network (FCNN) – face major challenges, including high dimensionality, large variable disparities, and limited data availability. Specifically, kriging models suffer from inefficient training processes, while FCNN models struggle with diminished prediction accuracy when confronted with diverse input sets. To address these challenges, this paper introduces two improved surrogate models by incorporating variable sensitivity into the kriging and FCNN models. They employ the analysis of variance to identify the global sensitivity of input variables and utilise K-means clustering to group variables based on their sensitivities. For the kriging model, auxiliary parameters corresponding to the number of clusters are introduced to replace hyperparameters, accelerating model training while maintaining high accuracy. For the FCNN model, input variables are grouped based on their sensitivities, with specialised expert networks handling each group, and a gating network combining their outputs to improve prediction accuracy. The effectiveness of these methods is demonstrated through numerical function examples and two aerodynamic data modelling scenarios: the FDL-5A hypersonic vehicle and the Saenger aerospace plane carrier wing. Results indicate that the proposed approaches significantly enhance the kriging model’s training efficiency, achieving a 98% reduction in hyperparameter tuning time compared to conventional method, with minimal sacrifice in accuracy. Simultaneously, the modifications to the FCNN model not only improve its prediction accuracy but also increase its overall practical utility in engineering applications.

Keywords:

Aerodynamic data modelling

data-driven surrogate model

kriging

fully connected neural network

variable sensitivity

Improved data-driven surrogate models by incorporating variable sensitivity for aerodynamic data modeling

Abstract

Journal

Organization