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

Hairun Xie, Jing Wang, Miao Zhang

TL;DR

This paper introduces a novel "knowledge-embedded meta-learning" model that combines theoretical aerodynamic principles with machine learning to predict airfoil lift coefficients. By integrating domain knowledge into the model architecture, it achieves better generalization and interpretability compared to traditional black-box AI models, offering a promising path towards more trustworthy AI in aerospace design.

What Problem Does This Paper Address?

Predicting the aerodynamic performance of airfoils, specifically their lift coefficient (CL), is crucial for aircraft design. Traditional methods like Computational Fluid Dynamics (CFD) are accurate but extremely time-consuming and computationally expensive, especially during the iterative optimization process. While machine learning (ML) models can speed this up, they often suffer from poor generalization to unseen airfoil designs and lack interpretability, making engineers hesitant to trust their predictions. This "black box" nature hinders their adoption in critical engineering applications. The paper aims to bridge this gap by developing an ML model that leverages existing theoretical knowledge about airfoil lift to improve both accuracy and trustworthiness.

Methods

The core innovation is a "knowledge-embedded meta-learning" model. This architecture consists of two main components:

  • Primary Network: This network learns the direct relationship between the angle of attack ($\alpha$) and the lift coefficient (CL). It's designed to capture the known theoretical behavior of the lift curve, which has a linear region at low angles of attack and a non-linear region at higher angles, transitioning at a critical angle ($\alpha_{crit}$). The network incorporates trainable parameters like $C_{L0}$ (lift coefficient at zero AoA) and the slope of the linear portion ($C_{L\alpha}$), which are physically meaningful.
  • Hyper Network: This network takes airfoil geometry (represented by a set of parameters) as input and predicts the parameters for the primary network. This is where the "meta-learning" aspect comes in – the hyper network learns how to "learn" the primary network's parameters for a given airfoil.

This architecture allows the model to embed theoretical knowledge about the lift curve into the primary network while using the hyper network to adapt these parameters based on specific airfoil geometries. The paper explores three variations of this model: a "Base Model," one with "+Weight Decay" to encourage sparsity, and one with "-Activation" to enforce a linear relationship between geometry and primary network parameters.

Datasets

The study utilizes a dataset of 2000 airfoils with their corresponding CL-$\alpha$ curves. These airfoils were sampled with constraints on leading-edge radius and maximum thickness to ensure diversity. The dataset was split into 80% for training and 20% for testing. For each airfoil, approximately 50 $\alpha$ values were sampled to generate the CL-$\alpha$ curves. The CFD simulations used to generate this data were performed using the Reynolds Average Navier-Stokes (RANS) equations with the shear stress transport (SST) turbulence model, solved with the CFL3D code.

Key Insights and Ideas

  • Knowledge Embedding Enhances Generalization: By explicitly incorporating theoretical knowledge about the lift curve into the model architecture, the proposed method significantly improves generalization capabilities compared to standard "black-box" neural networks.
  • Interpretability is Achievable: The knowledge-embedded structure, particularly the physically meaningful parameters in the primary network, allows for interpretable analysis. This helps understand how airfoil geometry influences aerodynamic characteristics.
  • Hyper Network as a Geometry-to-Physics Mapper: The hyper network acts as a crucial bridge, translating geometric features into the parameters that govern the airfoil's aerodynamic behavior.
  • Dual-Driven Approach: The model's success stems from a "dual-driven" approach, combining the strengths of data-driven learning with the robustness of domain knowledge.
  • Identifying Outliers: The interpretability tools (Integrated Gradients and Saliency) helped identify specific airfoil geometries (like sample-g) that exhibit complex flow phenomena (double shock waves) leading to prediction challenges.

Key Results

  • Accuracy: The "Base Model" achieved the lowest Mean Absolute Error (MAE) of $1.517 \times 10^{-3}$ and a competitive Mean Absolute Error (MAXE) of $5.299 \times 10^{-2}$ on the test set, outperforming the "+Weight Decay" and "-Activation" variants.
  • Generalization: Compared to a standard Artificial Neural Network (ANN), the proposed models showed much smaller discrepancies between training and testing errors, indicating superior generalization. For instance, the testing MAE for the ANN was more than double its training MAE, while the proposed models showed near-parity.
  • Interpretability: Analysis using Integrated Gradients and Saliency methods revealed that the latter portion of the lower airfoil surface has the greatest positive influence on the lift coefficient ($C_{L0}$). This provides actionable insights for designers.
  • Handling Complex Flows: The study identified that airfoils exhibiting double shock waves (like sample-g) pose challenges for the model, highlighting the limitations when knowledge doesn't fully capture complex physical phenomena. Excluding such samples significantly improved the model's performance.

Limitations and Future Work

  • Knowledge as a Restriction: The paper acknowledges that while knowledge embedding is beneficial, it can also act as a restriction if the embedded knowledge doesn't perfectly align with the data or if it fails to capture complex phenomena like double shock waves.
  • Outlier Handling: The model struggles with outliers exhibiting complex flow physics not fully captured by the embedded theoretical knowledge. Future work could focus on better handling these cases or developing more robust theoretical frameworks for such phenomena.
  • Discovering Universal Knowledge: The authors suggest that future research should explore how to use data more effectively to discover universal knowledge, rather than just embedding existing theories.
  • Further Investigation of Double Shock Waves: The exceptions caused by double shock waves warrant further investigation to develop theories and models that can accurately predict and handle these challenging flow patterns.

Summary updated

Abstract

Aerodynamic performance evaluation is an important part of the aircraft aerodynamic design optimization process; however, traditional methods are costly and time-consuming. Despite the fact that various machine learning methods can achieve high accuracy, their application in engineering is still difficult due to their poor generalization performance and "black box" nature. In this paper, a knowledge-embedded meta learning model, which fully integrates data with the theoretical knowledge of the lift curve, is developed to obtain the lift coefficients of an arbitrary supercritical airfoil under various angle of attacks. In the proposed model, a primary network is responsible for representing the relationship between the lift and angle of attack, while the geometry information is encoded into a hyper network to predict the unknown parameters involved in the primary network. Specifically, three models with different architectures are trained to provide various interpretations. Compared to the ordinary neural network, our proposed model can exhibit better generalization capability with competitive prediction accuracy. Afterward, interpretable analysis is performed based on the Integrated Gradients and Saliency methods. Results show that the proposed model can tend to assess the influence of airfoil geometry to the physical characteristics. Furthermore, the exceptions and shortcomings caused by the proposed model are analysed and discussed in detail.
BibTeX
@inproceedings{xie2023knowledge,
  author = {Hairun Xie and Jing Wang and Miao Zhang},
  title = {{Knowledge-embedded meta-learning model for lift coefficient prediction of airfoils}},
  booktitle = {Expert},
  year = {2023},
}
Figure