Knowledge-embedded meta-learning model for lift coefficient prediction of airfoils

The evaluation of aerodynamic performance is a crucial aspect of aircraft design optimization. However, traditional methods are time-consuming and expensive. While machine learning techniques can achieve high accuracy, their application in engineering is challenging due to their poor generalization and ”black box” nature. In this paper, a knowledge-embedded meta-learning model is proposed that integrates data with theoretical knowledge to predict lift coefficients for a supercritical airfoil at various angles of attack. The model employs a primary network to capture the relationship between lift and angle of attack, and a hyper network to encode geometry information and predict unknown parameters in the primary network. Three models are trained with different architectures to provide various interpretations. The model exhibits better generalization capabilities and competitive prediction accuracy compared to conventional neural networks. Additionally, interpretable analysis is performed using Integrated Gradients and Saliency methods, revealing that this model can assess the influence of airfoil geometry on physical characteristics. Furthermore, exceptional cases are investigated where the model’s performance is reduced, analyzing the impact of knowledge limitations on a knowledge-embedded predictive model.