5TH EUROPEAN CONFERENCE FOR AERONAUTICS AND SPACE SCIENCES (EUCASS)

Applying surrogate modelling for dynamic stall simulation

Vadim Surpin*, Alexander Bernstein**, Yuri Sviridenko***

* Institute for Information Transmission Problems RAS, Moscow, Russia
[email protected]
**International Research Institute for Advanced Systems, Moscow, Russia
[email protected]
***Central Aerohydrodynamic Institute, Moscow, Russia
[email protected]

Abstract
Presently, numerical techniques based on Navier-Stokes equation to investigate a dynamic stall event
are widely used. Application of CFD in design tasks and dynamic real-time simulation of objects is
limited by extensive usage of computing resources. The constructed, based on data surrogate model
performs calculations in less than one second that enables its application to early stage design tasks
and nearly real-time simulation tasks.

1. Introduction

Presently, numerical techniques based on Navier-Stokes equation to investigate a dynamic stall event are widely used
[1, 2]. Applying numerical methods to solve these equations allows defining the pressure distribution as well as
forces and moments affecting the aerodynamic airfoil in dynamic stall mode. Despite the significant progress in
computing and numerical techniques in Computational Fluid Dynamics (CFD) software, application of CFD codes
for calculation viscous unsteady flow of aerodynamic units requires extensive calculating resources. It does not allow
using CFD effectively in designing tasks and dynamic real-time simulation of objects (e.g. flight test-rigs).

The Predictive Simulation Technology [3] based on Data Handling developed in IRIAS&IITP RAS is used in the
research for constructing the fast surrogate models (or metamodels) for predicting the integral and distributed
aerodynamic time characteristics of the airfoils under the unsteady flow condition with dynamic stall for a given
airfoil geometry, unperturbed flow parameters and parameters describing time fluctuation of airfoils Angle of
Attack in time.

The constructed surrogate model is based on the data the results of the computational experiments performed by
the in-house CFD-solver FlowVision intended for modeling 3D laminar and turbulent steady and unsteady gas and
liquid flows in complex geometries.

2. Problem statement

The problem of predicting the integral and distributed aerodynamic time characteristics of airfoils under the unsteady
flow condition with dynamic stall is likely to arise in the industrial test cases. Thus it became the subject of an initial
study for the construction of a Fast Data-based Reduced Order Model on the basis of the results of the computational
experiments with chosen CFD solver considered as Full Order Model.

The dynamic stall event should be essentially considered under various industrial tasks and following conditions:
passenger aircraft penetration in high turbulence zones, light aircraft maneuvers, helicopter rotor blade flow under

The research leading to these results has received funding from the European Communitys Seventh Framework Programme
(FP7 / 2007-2013) under a grant agreement number 233665. FFAST (Future Fast Aeroelastic Simulation Technologies) is a
collaborative research project aimed at developing, implementing and assessing a range of numerical simulation technologies to
accelerate future aircraft design. Advances in critical load identification and reduced order modelling methods will potentially
provide a step change in the efficiency and accuracy of the dynamic aeroelastic loads process. The partners in FFAST are:
University of Bristol, INRIA, CSIR, TU Delft, DLR, IRIAS, University of Liverpool, Politecnico di Milano, NUMECA,
Optimad Engineering, Airbus, EADS-MS, IITP and UCT.

Copyright 2013 by Vadim Surpin, Alexander Bernstein, Yuri Sviridenko. Published by the EUCASS association with
permission.
Vadim Surpin, Alexander Bernstein, Yuri Sviridenko

various regimes and so on. For instance, one of the rotor blade aerodynamic characteristic features is that rotor blade
cross-sections work under the unsteady flow at almost all helicopter flight regimes [4].

Dynamic stall simulation is a complicated problem and the main investigation method employed to date is carrying
out extremely expensive experiments in wind tunnels [5]. Moreover, airfoil characteristics, which are traditionally
used, were obtained from wind tunnels during steady blowing to calculate forces and moments of helicopter rotors
[5]. For new helicopter airfoils such experimental data are not available. Due to the above difficulties, mathematical
simulation under dynamic conditions is getting to be the main tool for the study of airfoil unsteady aerodynamic
characteristics [6]. However, despite the significant progress in computing and numerical techniques, detailed CFD
codes being applied to calculate detached viscous unsteady flow requires extensive computational resources, which
limits its usefulness in real-time design and simulating object dynamics (e.g. flight test-rigs). For the above reasons,
generation of fast Reduced Order Model for the time-consuming CFD-code, which describes loading of aircraft wing
cross-sections or helicopter rotor blade under dynamic stall, was selected as a test case for the research.

Meanwhile the surrogate model should correctly indicate influence of the main factors: Mach and Reynolds numbers
of incoming flow, the law of variation of angle of attack, cross-section geometry on unsteady forces and moments
affecting the airfoil as well as on pressure distribution along the wing airfoil or blade section.

3. Experiment setup

To model of dynamic stall under unsteady flow the following experiment proposed. The airfoil performs a pitching
motion is expressed as

( ) ( ( ) ) (1)

where: ( ) represents the instantaneous angle of attack, initial angle of attack, amplitude of airfoils
oscillation, Strouhal number, flow velocity, length of airfoils chord.

3.2 Input data

The input data required for a detailed unsteady CFD analysis consists of:
Description of airfoil geometry;
Unperturbed flow parameters: Mach number ( ) and Reynolds number ( ) of the incoming flow;
Parameters ( ) describing a fluctuation of airfoils Angle of Attack in time - the current angle
of attack ( ).

3.3 Output Data

The output data consist of two functions of time :

The forces and moments affecting the airfoil;
Pressure distribution along the airfoil.

3.4 CFD-solver

For solving the discussed problem, the CFD code FlowVision has been chosen as Full Order Model a source of the
experimental data for constructing the surrogate models. FlowVision is a general purpose CFD software for
modeling 3D laminar and turbulent steady and unsteady gas and liquid flows in complex geometries.

3.4 Settings

To get results that can be used for surrogate model training the number of time steps of the full order CFD
experiment was set to a constant value so that every experiment produced exactly two periods of airfoil
oscillations and every period contained fixed number of data points . Every data point contains lift coefficient
value for the full airfoil and pressure distribution over both sides of the airfoil.

2
 INSTRUCTIONS FOR THE PREPARATION OF PAPERS

The data acquired for the first period is transient, i.e. the first period is the time required for the CFD code to set up
and it isnt used to surrogate model training. Data of the second and further periods considered to be valid.

4. Experiment results

An overview of the physical processes of dynamic stall is shown on Fig. 1. It is a velocity vector field which shows a
complex structure of air flows near the airfoil in presence of dynamic stall effect. The dynamic stall effect

Figure 1: Flow speed vector field representing the dynamic stall effect.

For the purposes of building surrogate model there were produced about 300 full order computational experiments. A
large part of them is fast experiments on a rough grid. A smaller part of experiments is higher accuracy experiments
(Fig. 2). Their results are used as a reference to ensure the lower accuracy results reflect the physics of the dynamic
stall effect.

Lower accuracy model reflects the general behavior of the flow but direct comparison to the high accuracy model
using standard techniques such as root mean square error gives large values for residuals due to phase effects.

Despite this fact further research will be based on rough models as the reflect the basic physical effects and fast
enough to produced required set of training data.

1,8
Higher
1,6 accuracy
computations
1,4 Faster
computations
1,2

0,8

0,6

0,4

0,2

0
10 12 14 16 18 20 22 24 26 28
Figure 2: High accuracy (light line) and fast (dark line) lift coefficient CFD computation results for input parameters
, and

3
Vadim Surpin, Alexander Bernstein, Yuri Sviridenko

4. Surrogate modelling technology

The Predictive Simulation Technology based on Data Handling developed in IRIAS&IITP RAS is used for
constructing the surrogate models (or metamodels) in this research. This technology was specifically dedicated to
EADS Business Units (AIRBUS, EUROCOPTER, ASTRIUM, etc.), with a first successful phase called MACROS
(Multidisciplinary Aeronautic Capability Research On Simulation) and completed in 2009. The technology [3] is
based on advanced mathematical core and allows integrating the domain-specific knowledge, models and data into
Surrogate models.

MACROS technology is implemented in MACROS Technological Tool for predictive modeling and simulation
based on data handling. The MACROS Technological Tool allows reducing the dimension of the input vectors,
constructing the data-based dependencies and evaluating its accuracy and doing the other data-processing procedures
in automatic mode for designing the surrogate models.

The main points of the technology:

(1) The considered problem is to predict the values of the characteristic Y of some chosen object for the
specified input vector X that includes the objects digital description and describes also its environment and
control parameters.
(2) Let be some chosen initial model (or method) considered as Full Order Model (FOM) that allows
obtaining the value (response, output) for the specified input vector . The model M determines the
response function

( ) (2)

where and are -dimensional and -dimensional Euclidean spaces respectively.

(3) Usually FOM is either a full-scale experiment or a computational experiment based on a solution of the
differential equations and hence may be expensive and time-consuming. The problem is to construct the
new Reduced Order Model (ROM) that is close to the FOM (it has the same accuracy) but essentially
increases the calculation speed.
(4) The results of the experiments with the initial model M are available. These results form the Data Set

{( ( )) } (3)

(5) The Learning Data Set is used for constructing the Data-based dependency

( ) ( | ) (4)

providing an approximate equality

( ) ( ) (5)

that has to hold for all and not only { }.

The High Dimension Approximator Generic Tool (a part of the MACROS Technological tool) is used for
constructing the Data-based dependency.

Thus the new Data-based model described by the constructed dependency ( ) can replace the initial
model and may be regarded as a Reduced Order Model, or a surrogate model, or a metamodel (model over
model).

4
 INSTRUCTIONS FOR THE PREPARATION OF PAPERS

5. Application of the technology to the dynamic stall approximation

To build a dynamic stall surrogate model based on the MACROS high dimension approximation tool we used
several approaches:

Direct MACROS approximator exploitation;

MACROS Approximator coupled with dimension reduction;

Approach based on time-series analysis.

First experiments using direct approximator exploitation approach were carried out using a training set of 50 data
vectors. Each data vector { } consists of data points where is a number of CFD solver
steps per one period.

Approximator build using showed ability to catch the tendency but the error value was large. Increasing the
number of experiments to 100 gave better results and further growth of leaded to error value decrease.

Along with extensive way of increasing number of data vectors used as a training base for the approximator, several
alternative approaches where taken to increase accuracy.

The first alternative is a Dimension Reduction approach. It is based on the idea that our data vector is quite large it
has at least three hundred sixty points used as a training data. There should be interdependency between points which
Dimension Reduction should eliminate. The approach give approximately the same error value as direct
approximator exploitation.

The second alternative approach is based on the fact that we know that interdependency exists. The data is a time-
series data and we can use the fact explicit. To employ the fact we used a time-series analysis methods.

The first attempt was to transform a data vector to a difference vector where each successive point is
a difference of original point value and previous point value. Index stands for number of time step or index of i-th
component of initial data vector . The approach gave the same results as the basic one.

Another attempt was to employ autoregressive methods to set up an explicit dependency between points of our time
series. In general, the approach states that each successive point is a function of previous points. The traditional
approach is to use linear function or apply some kernel function. Our approach is to use MACROS as a function
approximator:

( ) (6)

where non-linear function and random error. In our case was produced by the MACROS approximator and
represented as executable computer code.

The approach demonstrated root mean square error in the range of 5-10 percent and calculation time less than one
second. Fig. 3 shows a typical approximation result. The vertical axis is a lift coefficient and the horizontal axis is
angle of attack. The dark line is the approximation result while the light one is full order CFD result. Its relative
standard deviation is nearly 6 percent which is typical for the approach.

5
Vadim Surpin, Alexander Bernstein, Yuri Sviridenko

Figure 3: Full order (light line) and surrogate (dark line) lift coefficient computation results for input parameters
, and

6. Conclusion

Proposed approach using time-series analysis methods for data preprocessing and MACROS approximator as
surrogate model production tool showed ability to approximate lift coefficient data in presence of dynamic stall
effect with standard deviation in range 5-10% and calculation time less than 1 second.

References

[1] A. Gonzales, X. Munduate (2007). Unsteady modeling of the oscillating S809 aerofoil and NREI phase VI
parked blade using the Beddoes-Leishman dynamic stall model // In: Journal of Physics: Conference Series, 75,
p. 1 8, IOP Publishing.
[2] A. Spentzos, G. Barakos, K. Badcock, B. Richardo (2006). Modeling three-dimensional dynamic stall of
helicopter blades using computational fluid dynamics and neural networks // Proceedings of the Institute of
Mechanical Engineers, vol. 220, Part G: J. Aerospace Engineering, p. 605 618
[3] A.P. Kuleshov, A.V. Bernstein (2009). Cognitive technologies in adaptive models of complex plants. // In:
Keynote papers of 13th IFAC Symposium on Information Control Problems in Manufacturing (INCOM'09),
June 3-5, Moscow, Russia. pp. 70 - 81.
[4] W.J. McCroskey, L.W. Carr, and K.W. McAlister (1976). Dynamic Stall Experiments on Oscillating Airfoils //
AIAA Journal 1976 vol.14 no.1, p. 57-63.
[5] G. Leishman (1990). Dynamic stall experiments on the NACA 23012 aerofoil, Experiments in Fluids //
Springer-Verlag.
[6] K. Nguyen, W. Johnson (1998). Evaluation of Dynamic Stall Models with UH-60A Airloads Flight Test Data //
AHS International 54th Annual Forum Proceedings, Washington, D.C., May 20-22, pp. 576587.