Geometric feature knowledge-driven surrogate-based optimization via aerodynamic supervised autoencoder

•An aerodynamic supervised autoencoder is proposed to learn geometric feature correlated with aerodynamic responses with limited aerodynamic data.•The correlation between geometric features and aerodynamic responses is utilized to guide the initial sampling toward regions near the optimum.•The Eucli...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Aerospace science and technology Jg. 168; S. 111028
Hauptverfasser: Ma, Long, Wu, Xiaojing, Zuo, Zijun, Zhang, Weiwei
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier Masson SAS 01.01.2026
Schlagworte:
Aerodynamic shape optimization
Autoencoder
Feature learning
Geometric knowledge
Intelligent sampling
Feature learning
Aerodynamic shape optimization
Autoencoder
Intelligent sampling
Geometric knowledge
ISSN:1270-9638
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:•An aerodynamic supervised autoencoder is proposed to learn geometric feature correlated with aerodynamic responses with limited aerodynamic data.•The correlation between geometric features and aerodynamic responses is utilized to guide the initial sampling toward regions near the optimum.•The Euclidean distance between the predicted solution and the current optimum in the feature space is used as a penalty term to enhance the effectiveness of infill sampling.•The proposed optimization framework improves optimization efficiency by approximately twofold while achieving superior aerodynamic performance. Machine learning provides a promising approach for aerodynamic design. However, how to effectively learn from geometric or aerodynamic data to improve design performance remains a challenge. This work presents a geometric feature knowledge-driven surrogate-based optimization framework to accelerate the design process by exploiting geometric data. An aerodynamic supervised autoencoder, which linearly embeds limited aerodynamic data into the latent layer, is proposed to learn geometric features correlated with aerodynamic responses. Based on the learned feature knowledge, two approaches are developed to enhance sample quality for surrogate modeling in aerodynamic optimization. First, a promising subspace is identified in the feature space based on the correlation to guide the initial sampling toward regions near the optimum. Then, the Euclidean distance between the predicted solution and the current optimum in the feature space is used as a penalty term to enhance the effectiveness of infill sampling. The proposed framework is validated through aerodynamic optimization of the RAE2822 airfoil and the ONERA M6 wing. Results demonstrate that the aerodynamic supervised autoencoder can extract geometric features that are correlated with various aerodynamic responses using limited aerodynamic data. Compared to frameworks without feature knowledge, the proposed optimization framework improves optimization efficiency by about twice while achieving superior aerodynamic performance.
ISSN:1270-9638
DOI:10.1016/j.ast.2025.111028