A machine learning methodology to predict the aerodynamic performances of an Active Flow Control system on a 2D airfoil

Aug 4, 2026

Upload by Skyler

Investigates a machine learning workflow to predict aerodynamic performances of an Active Flow Control device, using a Sweeping Jet applied at the rudder hinge on a 2D vertical-tail airfoil (NACA 0012). The flow field is generated via CFD at Reynolds number Re=15 million and Mach number M=0.15, spanning sideslip angles β=0°, 5°, 10° and rudder deflections δ=0°, 10°, 20°. Jet velocity is represented with a square-wave signal varying frequency (50–150 Hz) and peak ejection velocity, while CFD coefficients train a one-hidden-layer feed-forward neural network (10 neurons) to output Cl, Cd, and Cm, achieving accurate prediction and about 20% sideforce enhancement under fixed f=100 Hz, Re=15 million, and M=0.15, with low computational cost after training.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages80

Reading Time3 h 22 min

AccessFree to download

Introduction

The oversized vertical tail surface problem

Aim of the thesis

Structure of the work

Theoretical Background

Flow equations

Neural networks

Neural Network training

Numerical Simulations

Flow conditions

2D Geometry

Domain and Boundary Conditions

A machine learning methodology to predict the aerodynamic performances of an Active Flow Control system on a 2D airfoil. TESI DI LAUREA MAGISTRALE IN AERONAUTICAL ENGINEERING Author: Carlo Cordoni Student ID: 976502 Advisor: Maurizio Boffadossi Co-advisor: Antonio Saporiti Academic Year: 2022-23 Abstract This thesis is the first step toward the design of a new active flow control device with the aid of Artificial Intelligence (AI) . A methodology was investigated for applying Machine Learning (ML) to predict the aerodynamic performances of a Sweeping Jet (SJ), which was applied to a 2D vertical tail section (NACA 0012) of an airplane, at the hinge of the rudder, to enhance the tail aerodynamic effectiveness. The flow field around the airfoil was studied with Computational Fluid Dynamics (CFD) at Reynolds’ number (Re) 15 million and Mach number (M) 0.15. The model was simulated at sideslip angles β of 0°, 5°, 10° and at rudder deflections δ of 0°, 10° and 20° . The SJ velocity was modelled by a square wave function of different frequencies f (50,100,150 Hz) and maximum ejected velocities 􀝒􀯆􀮺􀯑 (0,10,25,50,75,100 m/s) . The SJ action was successful in avoiding separation on the movable surface, provided sufficient momentum was injected on it. CFD aerodynamic coefficients were exploited to train a feed-forward neural network of one level of 10 neurons, 4 inputs (β, δ, 􀝒􀯆􀮺􀯑 ,f) and 3 outputs (Cl,Cd,Cm) . The network outputs were able to correctly predict the target CFD outputs. Moreover, the Neural Network was employed to obtain a sideforce enhancement of 20%, for β from 0° to 10° and for δ from 10° to 20°, by keeping fixed f (100 Hz), Re (15 million) and M (0.15). The successful demonstration of this machine learning methodology was confirmed by the small amount of time and by the low computational effort in providing aerodynamic coefficients, once the network was properly trained. Key-words: Artificial Intelligence, Machine-Learning, Neural-Network, Active-Flow-Control, Sweeping-Jet, Fluidic-Oscillator, Computational-FluidDynamics, Tail-Vertical-Surface Abstract in italiano Questo lavoro di tesi è il passo iniziale verso la progettazione di un nuovo dispositivo per controllare attivamente il flusso con il supportodell’intelligenza artificiale. È stata esplorata una metodologia per applicare il machine learning alla previsione delle prestazioni aerodinamiche di un Getto Oscillante (GO) nello spazio, applicato a una sezione 2D (NACA 0012) di una superficie verticale di coda, alla cerniera del timone, in modo da aumentarel’efficacia aerodinamica della superficie di coda. Il flusso d’aria attorno al profilo è stato studiato con la Fluidodinamica-Computazionale (CFD) anumeri di Reynolds (Re) di 15 milioni e numeri di Mach (M) di 0.15. Il modello è stato simulato ad angoli di incidenza laterale β di 0°, 5°, 10° e a deflessionidel timone δ di 0°, 10° e 20° . Le velocità del GO sono state modellate da onde quadra di diverse frequenze f (50,100,150 Hz) e velocità massime di efflusso 􀝒􀯆􀮺􀯑 (0,10,25,50,75,100 m/s) . La azione del GO si è dimostrata efficace nel prevenire la separazione sulla superfice mobile, a condizione di fornire sufficiente quantità di moto al flusso su di essa. I coefficienti aerodinamicicalcolati dalla CFD sono stati impiegati nel training di una rete neurale feedforward di un livello di 10 neuroni, 4 inputs (β, δ, 􀝒􀯆􀮺􀯑 ,f) e 3 outputs (Cl,Cd,Cm) . Gli outputs della rete hanno predetto correttamente gli outputs target calcolati con la. Inoltre, la Rete Neurale è stata sfruttata per ottenere un incremento della forza laterale pari al 20%, per valori di β da 0° a 10° e di δ da 10° a 20°, mantenendo fissati i valori di f (100 Hz), Re (15 milioni) and M (0.15) . Il buon esito e l’efficacia di questa metodologia basata sul machine learning sono stati confermati dalla ridotta quantità di tempo e dal basso costo computazionale con cui sono stati forniti i coefficienti aerodinamici, una volta che la rete

FAQs

How is the Active Flow Control system modeled for prediction?

A Sweeping Jet is applied at the rudder hinge of a 2D vertical-tail airfoil (NACA 0012), and CFD generates the aerodynamic data used to train the machine learning model.

Which inputs and outputs does the neural network use?

The network uses four inputs: sideslip angle β, rudder deflection δ, jet peak ejection velocity, and sweeping frequency f; it predicts three aerodynamic coefficients: Cl, Cd, and Cm.

What performance improvement is reported after training?

The trained network enables estimation of sideforce enhancement of about 20% for β from 0° to 10° and δ from 10° to 20%, while keeping f=100 Hz, Re=15 million, and M=0.15.

Aug 4, 2026

Upload by Skyler