2.0 REVIEW OF THE CRIAQ MDO 505 PROJECT CONCEPT

The CRIAQ MDO 505 project continues the CRIAQ 7.1 project's adaptive upper-surface wing concept and adds a real wing-tip structure, structural optimisation, new aerodynamic optimisation constraints, new control challenges, electrical actuation system, and classic and adaptive ailerons.

The full-scale wing model is an optimised structure with a 1.5m span and a 1.5m root chord, including its aileron, with a taper ratio of 0.72 and a leading-edge sweep of 8°.

The wing model box and internal structure is manufactured from aluminium with the composite adaptive upper surface extending from 20% to 65% of the wing model chord. The adaptive upper surface (skin) is specifically designed and optimised for this project as a carbon composite skin. The actuators are also specifically designed and manufactured to the project requirements. Four electric actuators are installed on two actuation lines, fixed to the centreribs and to the composite skin. Each actuator is capable of independent action. On each line the actuators are situated at 32% and 48% of the chord. These actuator positions were selected after analysing several other possibilities. From the analysis it was observed that positioning the actuators at equal distances from the ends of the composite skin returned the best results in term of aerofoil optimisation. These actuator positions also represent the optimisation points used during aerodynamic optimisation. The aileron's articulation is situated at 72% of the chord. Figure 1 presents a sketch of the wing model concept.

Figure 1.

CRIAQ MDO 505 wing-tip concept sketch.

3.0 BASE AEROFOIL PERFORMANCES

The aerofoil (base aerofoil), on which the wing model is based, was provided by one of the industrial partners and is a modified version of a supercritical aerofoil. Figure 2 presents the base aerofoil in non-dimensional coordinates.

Figure 2.

Base aerofoil.

The base aerofoil performance was evaluated with XFoil 6.96, an aerodynamic solver that is used for all aerofoil analyses throughout this paper. XFoil was developed by Drela and Youngren^{(Reference Drela and Youngren35)} and was chosen to be used in this paper because of its precision, effectiveness and rapid convergence. It is a solver that permits both inviscid and viscous analyses as well as geometrical modification of the aerofoil. The inviscid calculation is performed with a linear vorticity stream function panel method^{(Reference Drela36)}, to which a Karman-Tsien compressibility correction is added, allowing for good subsonic flow prediction. For its viscous calculations a two-equation lagged dissipation integral boundary layer formulation^{(Reference Drela37)} is used, which incorporates the eⁿ transition criterion^{(Reference Drela38)}. The flow in the boundary layer and in the wake interacts with the inviscid potential flow by means of a surface transpiration model.

Figures 3 presents the pressure distribution results for the base aerofoil, for an angle-of-attack range between –3° and 3° at Mach 0.15 and a Reynolds number of 2.15E+06.

Figure 3.

Pressure distributions for Reynolds number of 2.15E+06.

4.0 OPTIMISATION ALGORITHM

An ‘in-house’ genetic algorithm software was developed to optimise the base aerofoil's drag and transition. The genetic algorithm was chosen because of its ability to give a great number of possible solutions and obtain a global optimum in each case.

A genetic algorithm is a meta-heuristic method inspired from the process of natural selection. It is an iterative method that necessitates a high number of individuals and several generations to achieve its convergence; both the number of individuals and the number of generations are problem-dependent.

In a simple version of a genetic algorithm, a first generation of individuals, represented by their genes, is created randomly. The genes that represent each of the individuals are chosen based on their abilities to change at each generation, towards attaining the proposed objective. The first random generation is evaluated, and then a fitness value is associated with each individual. The fitness value is based on the results obtained from comparison with the objective requirements. Random individuals are then chosen as parents based on their fitness values. Through a simple mixing of their genes (with the simplest method being that of associating 50% of the genes from each parent to the children), new individuals are created. The sum of the new individuals forms a new generation. The process is repeated until a certain number of generations has passed or until the fitness value has reached a user-imposed limit^{(Reference Mitchell39-Reference Whitley41)}.

In the ‘in-house’ version of the algorithm, the individuals are represented by aerofoils and the genes that characterise them are the vertical displacements of the actuators situated at 32% and 48% of the chord. The fitness value of each individual aerofoil is calculated from a fitness function based on its aerodynamic performances. A first version of this ‘in-house’ genetic algorithm was used in Sugar et al^{(Reference Sugar Gabor, Koreanschi and Botez42)}.

Figure 4 presents the logic flow diagram of the genetic algorithm used for the aerofoil optimisation.

Figure 4.

Logic flow diagram of the genetic algorithm coupled with XFoil.

The aerodynamic performances are evaluated with XFoil 6.96 and are represented by the lift coefficient, C_L, the drag coefficient, C_D, the transition position and the skin friction coefficient variation with the chord, C_f. Based on these characteristics and the behaviour to be improved, single objective functions were developed and grouped in a fitness function to allow the selection of more than one objective if desired. The functions were developed on the most desired objectives encountered for aerofoil optimisation, for example: delaying transition for a more laminar flow on the upper surface, minimisation of the drag coefficient or maximising or minimising the lift in case of optimisation of a multi element aerofoil:

(1)$$\begin{equation} \begin{array}{l} {F_f} &=& {w_1}({C_{l\_m}} - {C_{l\_o}}) + {w_2}\frac{{{C_{l\_m}}}}{{{C_{l\_o}}}} + {w_3}{C_l} + {w_4}\frac{{{C_l}}}{{{C_d}}}\\ &&+{w_5}\frac{1}{{{C_d}^2}} + {w_6}\frac{{Tr}}{{{C_d}}} + {w_7}Tr + {w_8}\frac{1}{{Tr}} + {w_9}{C_f},\end{array} \end{equation}$$

where the values of w_i represent the weight associated with the searched-for aerodynamic characteristic. The weights are chosen by the user based on the importance attached to each aerodynamic characteristic it wants to optimise. The weights can have positive or negative values, but the sum of all the weights in the fitness function should not exceed the absolute value of 100, which corresponds to attaching 100% importance to an objective in detriment of the others.

Although some of the objective functions might seem to be redundant, actually they explore different behaviour combinations. For example, if a double objective of maximising lift and minimising drag is searched, one can either work with two objective functions, giving them weight based on the importance attached to both lift and drag, or the user could choose one objective function that contains both terms, thus giving them equal value or letting the optimisation program to find the best combination of lift and drag.

To analyse, XFoil needs the aerofoil coordinates, but the optimisation algorithm only returns the vertical displacements of the points where the actuators are situated. As such, a reconstruction method that enforces a tangency condition as well as an iso arc-length condition is necessary for the upper surface of the aerofoil (between 20% and 65% of the chord, corresponding to the flexible skin); the other coordinates are kept identical to those of the base aerofoil. The reconstruction is incorporated in the optimisation algorithm.

Several different parameterisation techniques have been developed and applied over the years for aerofoil design, such as the polynomial representation^{(Reference Abbott and Doenhof43)}, the CST method created by Kulfan^{(Reference Kulfan and Bussoletti44)}, and Non-Uniform Rational B-Splines^{(Reference Berbente, Mitran and Zancu46)}. However, these methods either parameterise the complete aerofoil and thus are not able to represent only a limited part of it, or they are difficult to implement when only the reconstruction of a specific part is required. They are also not judged sufficiently time-efficient to solve the problem presented here. Therefore, spline functions^{(47,Reference Herrera, Lozano and Verdegay48)} were selected for the reconstruction of the upper surface of the wing model aerofoil.

The best-known spline interpolation is the ‘cubic spline’, which ensures continuity up to and including second-order derivatives and allows for the calculation of the curvature radius. Finch and Friswell^{(Reference Fincham and Friswell49)} use a cubic spline interpolation for their morphing trailing-edge system.

The cubic spline is represented by the third-degree polynomial:

(2)$$\begin{equation} {P_{3,i}}(x) = {y_i} + {m_i}(x - {x_i}) + {b_i}{(x - {x_i})^2} + {a_i}{(x - {x_i})^3}, \end{equation}$$

which describes the behaviour of the splines at each interval h_i (Equation 4).

The parameters a_i and b_i are functions of the slope m_i calculated in each node i. The slope is the solution to the tri-diagonal linear system:

(3)$$\begin{equation} {\rho _i}{m_{i - 1}} + 2{m_i} + {\lambda _i}{m_{i + 1}} = {d_i},{\rm{ }}i \in \overline {2,(N - 2)} , \end{equation}$$

where

(4)$$\begin{equation} \begin{array}{l} {\rho _i} = \frac{\displaystyle{{h_i}}}{\displaystyle{{h_{i - 1}} + {h_i}}},{\rm{ }}{\lambda _i} = 1 - {\rho _i},{\rm{ }}{d_i} = \frac{\displaystyle{3{\lambda _i}({y_{i + 1}} - {y_i})}}{\displaystyle{{h_i}}} + \frac{\displaystyle{{\rho _i}({y_i} - {y_{i - 1}})}}{\displaystyle{{h_{i - 1}}}}\\ {h_i} = {x_{i + 1}} - {x_i}; \end{array} \end{equation}$$

to which we add relations that replace the continuity conditions for the first and second derivatives that cannot be imposed on x_N. These conditions could either be imposed as values for the end slopes m₁ and m_{N ,} or they could be given in a more general form, through relations with their neighbouring slopes.

For our problem, we have chosen to add the continuity conditions in this more general form through relations with their neighbouring slopes using a particular case of the cubic spline interpolation, namely, the ‘natural cubic spline interpolation’, and which is defined by the following conditions at the ends x_i:

(5)$$\begin{equation} {P^{\prime\prime}}({x_1}) = {P^{\prime\prime}}({x_N}) = 0 \end{equation}$$

By replacing Equation (5) in Equation (3) we obtain the following linear system for the end slopes m₁ and m_N:

(6)$$\begin{equation} \left\{\begin{array}{l} 2{m_1} + {m_2} = 3\frac{\displaystyle{{y_2} - {y_1}}}{\displaystyle{{h_1}}}\\[6pt] {m_{N - 1}} + {m_N} = 3\frac{\displaystyle{{y_N} - {y_{N - 1}}}}{\displaystyle{{h_{N - 1}}}} \end{array} \right. \end{equation}$$

By imposing these conditions, the following integral is minimised:

(7)$$\begin{equation} I = \int\limits_{{x_1}}^{{x_N}} {{{[{f^{\prime\prime}}(x)]}^2}dx} , \end{equation}$$

where f(x) is the exact function that is approximated by the spline interpolation. Minimising the above integral by imposing the natural conditions (Equation 5) produces the smoothest cubic spline interpolation; therefore, this type of interpolation is chosen to reconstruct the aerofoil shapes.

After the reconstruction of the aerofoils based on the vertical displacements of the actuation points and analysed with XFoil, the aerofoils are evaluated with the fitness function. Based on their results, they are graded from 1 to 10, where 10 is the grade given to the best aerofoil.

The next generation of aerofoils is created from the present one, with each aerofoil in the current generation having at least one chance at being selected as parent.

To ensure that the choice of the parents is random, and thus to give the most chances to those aerofoils with higher grades, a probability function was created, which returns values between 1 and 10.

(8)$$\begin{equation} {P_S} = 11 - \max (\min ({({\mathop{\rm int}} (random(0) \cdot {10^{\frac{1}{{{A_t}}}}}))^{{A_t}}},10),1) \end{equation}$$

The obtained value is then compared with the grades of each aerofoil, and those grades that match it are grouped. From this group one aerofoil is randomly chosen as ‘parent’. The process is repeated for a number of times that is equal to the number of parents used to create the new aerofoil. In our case, we used the classical approach of one ‘mother’ and one ‘father’.

When all the parents are chosen, they are passed through the cross-over and mutation processes^{(Reference Deb and Agrawal50)}. The ‘cross-over routine’ used has a two-steps approach. For the first 10 generations, the genes of the parents are mixed in equal proportions. This first step hastens the convergence process and leads the optimisation towards the global optimum area. The second step, applied for the remaining number of generations, is a cross-over function derived from a simulated binary cross-over technique^{(Reference Deperrois51)} coupled with a ‘random number generator’ function.

(9)$$\begin{equation} \begin{array}{l} random(0)\\ if(random(0) > = 0.5)then\\ child = 0.5 \cdot (1 + random(0)) \cdot parent1\\ else\\ child = 0.5 \cdot (1 + random(0)) \cdot parent2\\ endif \end{array} \end{equation}$$

This second step is used to pinpoint, as accurately as possible, the best solution from the multitude of solutions found in the global optimum area.

Each new aerofoil in each generation is passed through a ‘mutation routine’ in which, based on the probability of occurrence (which is a value introduced by the user – 0.1% from the individuals in a generation in our case), is affected by gene mutation.

The mutation process, if it occurs, depends on the amplitude of mutation (also user-dependent), which in turn depends on the value to be mutated. The amplitude of mutation is usually a small percentage (2% in this case) of the gene's values derived after cross-over.

From mutation, the new aerofoils are passed through a verification process. Here, are verified the conditions related to actuator's maximum and minimum displacements and delta values between the actuator displacements. The conditions are derived from multidisciplinary work, as they are provided from aerodynamic, structural and control calculations and limitations.

If an aerofoil fails the requirements, the selection, cross-over and mutation processes are reiterated until a valid aerofoil is obtained. If a certain number of iterations (10,000 is the value imposed for this specific problem) have passed without yielding valid results, the optimisation process stops.

Under normal conditions, the processes of selection, cross-over, mutation and verification are repeated for each pair of parents until a fixed number of aerofoils in a generation is reached. This number of aerofoils (also referred as generation) is set by the user at the beginning of the optimisation process.

The complete process of reconstruction, analysis and evaluation is repeated until the maximum number of generation is reached.

To improve the overall convergence of the optimisation process a tournament is introduced. A tournament takes place after the current generation is analysed and graded and before the creation of the new generation. The tournament compares the worst aerofoils from the current generation with the best ones from the previous generation and replaces the former with some of the latter, thus favouring the propagation of good genes from the older generation to the current and then on to the future generations.

This tournament process hastens the convergence of the optimisation and for our problem of optimising a specific part of the aerofoil it reduces the number of generations from 40 to 20 and the number of aerofoils in a generation from 50 to 40.

6.0 AEROFOIL OPTIMISATION IMPACT ON WING MODEL PERFORMANCES

To fully understand the impact of the results obtained on the aerofoil optimisation, an analysis of the wing model, with its geometry based on the optimised aerofoils, is done using the XFLR5 codeReference Deperrois⁵¹. XFLR5 is an analysis tool for aerofoils, wings and aircrafts operating at low Reynolds numbers. It includes XFoil's direct and inverse analysis capabilities, as well as wing design and analysis capabilities based on Lifting Line Theory^{(Reference Sivells and Neely52)}, the Vortex Lattice Method^{(Reference Maskew53)} and the 3D Panel Method^{(Reference Katz and Plotkin54)}.

For the XFLR5 analysis of a wing, there are three steps to be followed:

1. 1. Analysis of the aerofoil(s) composing the wing using a multi-threaded batch analysis, which allows the analysis of multiple aerofoils at a specific speed over a range of Reynolds numbers and ranges of angles of attack using XFLR5’s XFoil section.

2. 2. Construction of the wing model, based on the aerofoil(s) analysed in the previous step. This step requires the number of sections (minimum of two root and tip sections), the span and chord dimensions for each section and, if present, the offset (m), dihedral and twist angles. Finally, the wing model needs the total number of panels required for the calculations in each direction for each section.

3. 3. Analysis of the wing model using one of the following methods: Lifting Line Theory, the Horse-shoe Vortex Lattice Method, the Ring Vortex Lattice Method or the 3D Panel Method.

For the present analysis, the aerofoils are the base aerofoil and the aerofoils resulted from the optimisation process for each case. The wing model is created from four sections: sections 1 and 4, representing the root and the tip of the wing model –the corresponding aerofoil is the base aerofoil; and sections 2 and 3, which represent the actuator lines in the span length – the aerofoils corresponding to them are the optimised aerofoils, specific for each studied flight case.

Figure 11 presents the wing model for one flight case as it was created using XFLR5.

Figure 11.

Wing model definition in XFLR5 for Mach 0.15 and angle-of-attack of 0.25°.

The analysis was done at the Mach number of 0.15 over the same range of global angles of attack as the optimised aerofoils, using the 3D Panels Method option for aerodynamic analysis. The 3D Panel Method was chosen because the other methods were considered as insufficiently accurate for the analysis. The Lifting Line Method works only for wings with aspect ratios greater than 4, while this wing model has an aspect ratio of 2.9. The Vortex Lattice Method reduces the body to a middle surface with zero thickness, which eliminates the notions of upper and lower surfaces and returns only the difference between upper- and lower-surface pressure coefficients. The 3D Panel Method takes into account the three-dimensional geometry surface and gives more detailed results for the studied geometry.

Figures 12–14 present the global reduction of the drag coefficient as indicated in the global C_D versus lift coefficient, viscous C_D versus lift coefficient, and inviscid C_D versus lift coefficient graphs, for the original and morphed (optimised) wing model.

(12)$$\begin{equation} {C_{D\_total}} = {C_{D\_viscous}} + {C_{D\_inviscid}} \end{equation}$$

Figure 12.

Wing model total drag coefficient versus lift coefficient for Mach 0.15.

Figure 13.

Wing model viscous drag coefficient versus lift coefficient for Mach 0.15.

Figure 14.

Wing model inviscid drag coefficient versus lift coefficient for Mach 0.15.

Table 2 shows the values of the global and viscous drag and the global drag reduction for Mach 0.15 and each angle-of-attack. The drag is presented in counts, where one drag count is equal to 10⁻⁴.

Table 2

Total wing model's drag coefficient improvement

Figures 15 to 17 present the span distribution of the profile drag for each case analysed, showing the difference between the original and the morphed wing models. The profile drag is the most affected by any modification in the aerofoil shape. Here, it is presented in counts, where one count represents 10⁻⁴.

Figure 15.

Profile shape drag versus wing model span – angles of attack –0.5° to 0.25°.

Figure 16.

Profile shape drag versus wing model span – angles of attack 0.5° to 1.25°.

Figure 17.

Profile shape drag versus wing model span – angle-of-attack 1.5°.

The figures presented above show that even though the morphing area is situated only in the space between the spars of the wing model and the displacements are quite small, an overall wing model drag improvement takes place for all of the studied cases.

Each case shows that the main reduction is concentrated in the region between the actuation lines, which are situated at 0.56m and 1.117m along the span as presented in section 2 of this paper.

An exception is the case corresponding to Mach 0.15 angle-of-attack a equal to 1° where a numerical error appears which affects the value of the drag coefficient in small measure. Most probably the reduction is less than 2%, but as the trend shown in Fig. 16 is similar to the others, it can be assumed that there is an approximately 1% reduction for it as well.