Rapid Airfoil Inverse Design Method with a Deep Neural Network and Hyperparameter Selection

• Published in: International journal of aeronautical and space sciences Vol. 24; no. 1; pp. 33 - 46
• Main Authors: Tyan, Maxim, Choi, Cheol-Kyun, Nguyen, Tuan Anh, Lee, Jae-Woo
• Format: Journal Article
• Language: English
• Published: Seoul The Korean Society for Aeronautical & Space Sciences (KSAS) 01.02.2023; 한국항공우주학회
• Subjects: Aerospace Technology and Astronautics; Engineering; Fluid- and Aerodynamics; Original Paper; 항공우주공학; Deep neural network; Airfoil design; Database enhancement; Inverse design; Artificial intelligence
• ISSN: 2093-274X; 2093-2480
• DOI: 10.1007/s42405-022-00507-x

A deep-learning based rapid engineering design (DL-RED) algorithm is developed and presented for solving engineering inverse design problems. The algorithm generates an inverse design deep neural network (IDNN) by performing a two-stage hyperparameter selection and iteratively enhancing training database. The initial hyperparameter selection defines the configuration of IDNN that increases the accuracy of inverse design and reduces the network training time. Then database enhancement loop refines the training database by iteratively adding new samples to initial database based on results of validation. After the training database is generated the final hyperparameter selection is performed. A construction of Airfoil inverse design neural network is demonstrated in this paper. The Airfoil-IDNN generates a NACA 4-series airfoil that corresponds to target aerodynamic parameters given as an input. During the database enhancement loop the validation mean squared error has reduced to 57.37%, and totally to 96.48% after the final hyperparameter selection. The accuracy of the Airfoil-IDNN was demonstrated using two case studies. The first case study shows how accurate the network can generate an airfoil with arbitrary input parameters. The second one shows whether the IDNN can generate existing airfoils by a set of target parameters.