Figure 1: Steady-state Mach contours inside diuser. Freestream Mach number is 2.2.

1
1.1

Supersonic Inlet Flow Example

Overview and motivation

This example considers unsteady ow through a supersonic diuser as shown in Figure 1. The diuser operates at a nominal Mach number of 2.2, however it is subject to perturbations in the incoming ow, which may be due (for example) to atmospheric variations. In nominal operation, there is a strong shock downstream of the diuser throat, as can be seen from the Mach contours plotted in Figure 1. Incoming disturbances can cause the shock to move forward towards the throat. When the shock sits at the throat, the inlet is unstable, since any disturbance that moves the shock slightly upstream will cause it to move forward rapidly, leading to unstart of the inlet. This is extremely undesirable, since unstart results in a large loss of thrust. In order to prevent unstart from occurring, one option is to actively control the position of the shock. This control may be eected through ow bleeding upstream of the diuser throat. In order to derive eective active control strategies, it is imperative to have low-order models which accurately capture the relevant dynamics.

1.2

Active ow control setup

Figure 2 presents the schematic of the actuation mechanism. Incoming ow with possible disturbances enters the inlet and is sensed using pressure sensors. The controller then adjusts the bleed upstream of the throat in order to control the position of the shock and to prevent it from moving upstream. In simulations, it is dicult to automatically determine the shock location. The average Mach number at the diuser throat provides an appropriate surrogate that can be easily computed. 1

Figure 2: Supersonic diuser active ow control problem setup.

There are several transfer functions of interest in this problem. The shock position will be controlled by monitoring the average Mach number at the diuser throat. The reduced-order model must capture the dynamics of this output in response to two inputs: the incoming ow disturbance and the bleed actuation. In addition, total pressure measurements at the diuser wall are used for sensing. The response of this output to the two inputs must also be captured.

1.3

CFD formulation

The unsteady, two-dimensional ow of an inviscid, compressible uid is governed by the Euler equations. The usual statements of mass, momentum, and energy can be written in integral form as t t t Q dV + E dV + dV + Q dA = 0 p dA = 0 p Q dA = 0, (1) (2) (3)

Q (Q dA) + H (Q dA) +

where , Q, H, E, and p denote density, ow velocity, total enthalpy, energy, and pressure, respectively. The CFD formulation for this problem uses a nite volume method and is described fully in Lassaux [1]. The unknown ow quantities used are the density, streamwise velocity component, normal velocity component, and enthalpy at each point in the computational grid. Note that the local ow 2

velocity components q and q are dened using a streamline computational grid that is computed for the steady-state solution. q is the projection of the ow velocity on the meanline direction of the grid cell, and q is the normal-to-meanline component. To simplify the implementation of the integral energy equation, total enthalpy is also used in place of energy. The vector of unknowns at each node i is therefore
x i = i , q i , qi , H i T

(4)

Two physically dierent kinds of boundary conditions exist: inow/outow conditions, and conditions applied at a solid wall. At a solid wall, the usual no-slip condition of zero normal ow velocity is easily applied as q = 0. In addition, we will allow for mass addition or removal (bleed) at various positions along the wall. The bleed condition is also easily specied. We set q = m , (5)

where m is the specied mass ux per unit length along the bleed slot. At inow boundaries, Riemann boundary conditions are used. For the diuser problem considered here, all inow boundaries are supersonic, and hence we impose inlet vorticity, entropy and Riemanns invariants. At the exit of the duct, we impose outlet pressure.

1.4

Linearized CFD Matrices

The two-dimensional integral Euler equations are linearized about the steadystate solution to obtain an unsteady system of the form E dx = Ax + Bu dt y = Cx (6)

The descriptor matrix E arises from the particular CFD formulation. In addition, the matrix E contains some zero rows that are due to implementation of boundary conditions. For the matrices given here, the CFD model has 3078 grid points and 11,730 unknowns. There are two inputs: u1 is the bleed actuation massow input and u2 is the incoming density disturbance input. There is a single output: the average Mach number at the inlet throat. The CFD matrices are stored in the Matrix Market format with sizes as follows, where n = 11, 730.

A : n n matrix B : n 2 matrix C : n 1 vector E : n n matrix

Model Reduction Results

Model reduction results for the linearized system are shown here for several dierent reduction methods. The proper orthogonal decomposition (POD) is applied using frequency domain samples to generate a snapshot ensemble as described in [2]. The Arnoldi method is used to generate a basis that spans the Krylov subspace. For the results shown here, only an expansion about zero frequency is considered. In [3], models are derived using multiple Krylov expansion points. Finally, Fourier model reduction (FMR) is applied. A description of this method can be found in [4]. Note that most of the results presented here are also described more fully in this reference. Reduction of the nonlinear system using a trajectory piecewise-linear modeling approach is described in [5]. The rst transfer function of interest is that between bleed actuation and average Mach number at the throat. Bleed occurs through small slots located on the lower wall between 46% and 49% of the inlet overall length. Frequencies of practical interest lie in the range f /f0 = 0 to f /f0 = 2, where f0 = a0 /h, a0 is the freestream speed of sound and h is the height of the diuser. Figure 3 shows the magnitude and phase of this transfer function as calculated by the CFD model and three reduced-order models each of size k = 10. The FMR model was calculated by using 201 Fourier coecients (requiring a single CFD matrix factorization) to construct a Hankel matrix as described in [4]. The parameter 0 is used to dene the bilinear transformation in the FMR method, and was set to 0 = 0.5 for the case shown in Figure 3. This 200th -order FMR system was then further reduced to ten states using explicit balanced truncation. The POD model was obtained by computing 41 snapshots at 21 equally-spaced frequencies from f /f0 = 0 to f /f0 = 2 (requiring 21 complex CFD matrix factorizations). The Arnoldibased model was obtained by computing the rst ten Arnoldi vectors about s = 0. It can be seen from Figure 3 that the FMR model matches the CFD results well over the entire frequency range plotted, with a small discrepancy 4

2.5 2 |G(jw)| 1.5 1 0.5 0 CFD POD Arn FMR

0.5

1.5 f/f0

2.5

3.5

2 phase(G(jw))

0 CFD POD Arn FMR 0 0.5 1 1.5 f/f0 2 2.5 3 3.5

Figure 3: Transfer function from bleed actuation to average throat Mach number for supersonic diuser. Results from CFD model (n = 11, 730) are compared to FMR, POD and Arnoldi models with k = 10 states. at higher frequencies. The Arnoldi model matches well for low frequencies, but shows considerable error for f /f0 > 1.3. The POD model has some undesirable oscillations at low frequencies, and strictly is only valid over the frequency range sampled in the snapshot ensemble (f /f0 < 2). The performance of the POD and Arnoldi models can be improved by increasing the size of the reduced-order models. Figure 4 shows the results using 30 Arnoldi vectors and 15 POD basis vectors. The agreement at low frequencies is now very good for all models, but the POD and Arnoldi models still show discrepancy at higher frequencies. The POD model could be further improved by including more snapshots in the ensemble; however, each additional frequency considered requires an nth -order complex matrix factorization. The Arnoldi model could be further improved by increasing the size of the basis; however, this was found to result in unstable reduced-order models. This result highlights one of the major problems with the commonly used POD and Krylov-based reduction techniques. Because no rigorous statement about model quality can be made, the reduction becomes an ad hoc process that requires trial and error to obtain accurate, stable reduced-order models. In particular, for the POD, the choice of snapshot ensemble is critical. FMR is also applied to the transfer function between an incoming den5

2 CFD POD k=15 Arn k=30 FMR k=10

1.5 |G(jw)|

0.5

0.5

1.5 f/f0

2.5

3.5

2 phase(G(jw))

0 CFD POD k=15 Arn k=30 FMR k=10 0 0.5 1 1.5 f/f0 2 2.5 3 3.5

Figure 4: Transfer function from bleed actuation to average throat Mach number for supersonic diuser. Results from CFD model (n = 11, 730) are compared to FMR with k = 10 states, POD with k = 15 states, and Arnoldi with k = 30 states. sity perturbation and the average Mach number at the diuser throat. This transfer function represents the dynamics of the disturbance to be controlled and is shown in Figure 5. As the gure shows, the dynamics contain a delay, and are thus more dicult for the reduced-order model to approximate. Satisfactory results for this case could not be obtained using POD or Arnoldibased models. Results are shown for FMR with m = 200 and 0 = 5, 10. With 0 = 5, the model has signicant error for frequencies above f /f0 = 2. Choosing a higher value of 0 improves the t, although some small discrepancy remains. These higher frequencies are unlikely to occur in typical atmospheric disturbances, however if they are thought to be important, the model could be further improved by either evaluating more Fourier coecients, or by choosing a higher value of 0 . The 0 = 10 model is further reduced via balanced truncation to a system with thirty states without a noticeable loss in accuracy.

2.5 2 1.5 real(G(jw)) 1 0.5 0 0.5 1 0 0.5 1 1.5 2 f/f0 2.5 3 3.5 4 CFD m=200, 0=5 m=200, 0=10 k=30, 0=10

1 0.5 imag(G(jw)) 0 0.5 1 1.5 2 0 0.5 1 1.5 2 f/f0 2.5 3 CFD m=200, 0=5 m=200, =10 0 k=30, 0=10 3.5 4

Figure 5: Transfer function from incoming density perturbation to average throat Mach number for supersonic diuser. Results from CFD model (n = 11, 730) are compared to 200th -order FMR models with 0 = 5, 10. The 0 = 10 model is further reduced to k = 30 via balanced truncation.

References
[1] G. Lassaux. High-Fidelity Reduced-Order Aerodynamic Models: Application to Active Control of Engine Inlets. Masters thesis, Dept. of Aeronautics and Astronautics, MIT, June 2002. [2] K. Willcox. Reduced-Order Aerodynamic Models for Aeroelastic Control of Turbomachines. PhD thesis, Dept. of Aeronautics and Astronautics, MIT, February 2000. [3] G. Lassaux and K Willcox. Model reduction for active control design using multiple-point Arnoldi methods. AIAA Paper 2003-0616, 2003. [4] K. Willcox and A. Megretski. Fourier Series for Accurate, Stable, Reduced-Order Models for Linear CFD Applications. AIAA Paper 20034235, presented at 16th AIAA Computational Fluid Dynamics Conference, Orlando, FL, to appear SIAM J. Scientic Computing, 2004. [5] D. Gratton and K Willcox. Reduced-order, trajectory piecewise-linear models for nonlinear computational uid dynamics. AIAA Paper 20047

2329, presented at 34th AIAA Fluid Dynamics Conference, Portland, Oregon, June 2004.