Large Wing Model

TL;DR

The paper introduces the Large Wing Model (LWM), a probabilistic machine learning approach that predicts pressure distributions and lift coefficients for finite wings using limited experimental data, incorporating uncertainty and physics-based priors.

Contribution

The LWM employs a modified deep kernel learning architecture with Gaussian Processes to accurately predict wing aerodynamics from small datasets, including uncertainty quantification and extrapolation capabilities.

Findings

Model achieves less than 1.7% error in lift coefficient predictions.

Demonstrates good extrapolation to new airfoil sections.

Effectively captures three-dimensional wing effects.

Abstract

Developing a generalized aerodynamics prediction machine learning model for finite wings with different airfoil sections is challenging due to the vast parameter space and a relative scarcity of available data. This paper presents the Large Wing Model (LWM), a probabilistic machine learning model designed to predict pressure coefficient ($C_p$) distributions using a small, strictly experimental data set. From its uncertainty-aware $C_p$ predictions, the sectional and total wing lift coefficients ($c_l$, $C_L$) and their confidence intervals are calculated. The LWM features a modified deep kernel learning architecture, building a Gaussian Process model in a 15-dimensional space formed by 14 latent variables and the wing spanwise dimension. It is trained on an open-source database of wind tunnel measurements developed for this work. The Bayesian approach ingests uncertainties associated with experimental measurements and data digitization into the model. The model demonstrates satisfactory extrapolation abilities, enabling predictions on wings with new airfoil sections via the physics-driven prior formed from two-dimensional $C_p$ predicted by the Large Airfoil Model. The model accuracy is assessed for three test cases, rectangular wings with varying airfoil sections and operating conditions. For all test cases, the model performed well; the error in $C_L$ did not exceed 1.7\%. The model effectively captures three-dimensional effects such as those induced by wing tip vortices. Furthermore, constraining the posterior predictive space based on known probabilistic descriptions of lift improves the accuracy of the $C_p$ predictions. As a computationally efficient wing $C_p$ prediction model, the LWM facilitates the rapid exploration of the wing design space.