betaFeedback

WorksMathematics

Ordinary differential equations for the adjoint Euler equations

Abstract

Ordinary Differential Equations are derived for the adjoint Euler equations firstly using the method of characteristics in 2D. For this system of partial-differential equations, the characteristic curves appear to be the streamtraces and the well-known C+ and C- curves of the theory applied to the flow. The differential equations satisfied along the streamtraces in 2D are then extended and demonstrated in 3D by linear combinations of the original adjoint equations. These findings extend their well-known counterparts for the direct system, and should serve analytical and possibly numerical studies of the perfect-flow model with respect to adjoint fields or sensitivity questions. Beside the analytical theory, the results are demonstrated by the numerical integration of the compatibility relationships for discrete 2D flow-fields and dual-consistent adjoint fields over a very fine grid about an airfoil.

Introduction

In 1988, Jameson derived the continuous adjoint equations associated with the 2D and 3D Euler equations using general curvilinear coordinates [1]. With this landmark article, the fluid dynamics and aeronautical communities became better aware of the potential of the adjoint approach for design, that is, the possibility to calculate gradient information at a cost scaling with the number of functions to be differentiated, independently of the number of design parameters. The equations in [1] appeared to be a natural starting point for local optimizations involving a large number of design variables by using adjoint gradients. However, in that setting, the flow and the dual fields had to be calculated over a structured mesh.

Nine years later, Anderson and Venkatakrishnan [2, 3] and also Giles and Pierce [4] derived the corresponding equations in Cartesian coordinates thus allowing the application of the continuous approach (sometimes referred to as the differentiate-then-discretize approach) on all types of meshes and, in particular, on unstructured meshes. For the sake of simplicity, we present here the two-dimensional case only in which the adjoint equations read

−AT∂ψ∂x−BT∂ψ∂y=0,in Ω the fluid domain(1)-A^{T}\frac{\partial\psi}{\partial x}-B^{T}\frac{\partial\psi}{\partial y}=0,\quad\text{in }\Omega\text{ the fluid domain} \tag*{(1)}

where AA and BB are the Jacobian matrices of the flux vectors FxF_x and FyF_y of the Euler equations in the xx and yy directions respectively:

Fx=(ρuρu2+pρuvρuH)Fy=(ρvρuvρv2+pρvH),F_x=\begin{pmatrix} \rho u\\ \rho u^2+p\\ \rho uv\\ \rho uH \end{pmatrix} \qquad F_y=\begin{pmatrix} \rho v\\ \rho uv\\ \rho v^2+p\\ \rho vH \end{pmatrix},

with ρ\rho the density, (u,v)(u,v) the velocity components, pp the static pressure and HH the total enthalpy. In the most common case where the quantity of interest (QoI) is a line integral along the solid wall Γw\Gamma_w, it can be shown easily that the adjoint wall boundary condition is well-posed provided that the function of interest depends only on the static pressure [3]. In the classical case where the functional output of interest is the force on Γw\Gamma_w projected in direction dˉ\bar{d}, J=∫Γwp(nˉ⋅dˉ)dsJ = \int_{\Gamma_w} p(\bar{n}\cdot\bar{d})ds, the wall boundary condition reads

nˉ⋅dˉ+ψ2nx+ψ3ny=0on Γw.(2)\bar{n}\cdot\bar{d}+\psi_2 n_x+\psi_3 n_y=0 \quad\text{on }\Gamma_w. \tag*{(2)}

For the farfield of an external flow, as well as for the inlet and outlet of an internal flow, the boundary conditions are derived from the theory of local one-dimensional characteristic decomposition [2, 5]. Here, the continuous adjoint Euler equations and the associated boundary conditions are abbreviated as (AE). Along with the growing use of the adjoint method for shape optimization, goal oriented mesh adaptation and also meta-modelling, stability or control, great effort is being devoted to gain understanding in the mathematical properties of the (AE) solutions. The main results are summarized here before discussing the characteristic relations for the (AE) system.

After the derivation of the (AE) equations, the first demonstrated property was also due to Giles and Pierce [4]: in the common case where the function of interest is an integral along the wall, the authors proved that the first and last components of the adjoint vector ψ\psi, associated with mass and energy conservation, satisfy ψ1=Hψ4\psi_1=H\psi_4.

Besides, the integration by parts yielding (1) is not valid in the entire fluid domain in the presence of flow discontinuities. After a series of works dealing with the quasi-1D Euler equations – see [5, 6] and references therein – Baeza et al. [6] presented the equations complementing (1) along a shock line [6] (denoted here Σ\Sigma as in the original reference). The new equations are derived by introducing a complementary set of Lagrange multipliers, multiplying the Rankine-Hugoniot conditions, viewed as constraints on Σ\Sigma. Finally, the continuity of the adjoint field ψ\psi along Σ\Sigma is established, although ∇ψ\nabla\psi may be discontinuous across Σ\Sigma, as well as ψ\psi over Γw∩Σ\Gamma_w\cap\Sigma, and a so-called internal boundary condition is derived:

(∂ψ∂t)⋅(Fxtx+Fyty)=0on Σ(3)\left(\frac{\partial\psi}{\partial t}\right)\cdot(F_x t_x+F_y t_y)=0 \quad\text{on }\Sigma \tag*{(3)}

with tt the unit vector tangent to Σ\Sigma

Coquel et al., Lozano and Renac [7, 8, 9] have derived additional relationships by using (3), the jump operator applied to (1) across Σ\Sigma and the Rankine-Hugoniot equations.

The fact that ψ1=Hψ4\psi_1=H\psi_4 can be proven simply by forming the linear combination of the first three lines of system (1) with coefficients (1,u/2,v/2)(1,u/2,v/2). This yields Uˉ⋅∇ψ1−HUˉ⋅∇ψ4=0\bar{U}\cdot\nabla\psi_1-H\bar{U}\cdot\nabla\psi_4=0 (with Uˉ=(u,v)\bar{U}=(u,v) the velocity vector). Note that this was also derived in [4] by an approach based on physical source terms, constituting an important analysis technique for the adjoint field of usual QoIs. In particular, this method proved to be very fruitful to identify the zones where numerical divergence of the adjoint vector is observed and mathematical divergence of the solutions of (AE) is suspected. For the sake of clarity and brevity, we restrict the present discussion to 2D flows about lifting airfoils, and to two of these zones, namely the stagnation streamline and the wall, and to the lift and drag as functions of interest.

More precisely, Giles and Pierce [4] introduced four physical punctual source terms (or Green’s functions in the classical mathematical vocabulary) denoted here δR1\delta R^1, δR2\delta R^2, δR3\delta R^3, δR4\delta R^4. These terms are added to the right hand-side of the linearised Euler equations and correspond respectively to (i) a mass source at fixed stagnation pressure p0p_0 and enthalpy HH; (ii) a normal force; (iii) a change in HH at fixed pp and p0p_0; and (iv) a change in p0p_0 at fixed pp and HH. They are linearly independent. (We refer to the original reference for the detailed expression of these source terms.) The resulting changes in the QoI JJ, δJl\delta J^l, can be expressed as the integral over the domain of ψδRl\psi\delta R^l that is, the value at the source location since δRl\delta R^l is a Green’s function. These source terms also admit a physical interpretation and their influence on the flow can be understood in terms of mechanical principles, and sometimes even quantified finally providing insight in the adjoint field [4].

It has been observed that the lift adjoint exhibits numerical divergence at the stagnation streamline and at the wall at subcritical flow conditions. Also the drag and lift adjoint of a transonic airfoil exhibit numerical divergence at the same locations if the foot of at least one shock wave is located strictly upwind the trailing edge – see [10, 11] and references therein. Reference [9] includes a careful verification of this physical perturbations approach applied to the discrete adjoint with a preliminary assessment of the consistency between the linear (discrete adjoint) and the non-linear (flow perturbation) evaluations of the δJl\delta J^l. After this verification step, the non-linear perturbed flow approach has been used (considering the physical source terms point of view prior to the classical adjoint) and it appeared that: (a) a δR4\delta R^4 is the only source term causing a numerical divergence of δJ\delta J in the vicinity of the wall and stagnation streamline ; (b) in transonic condition, the numerical divergence of δCLp4\delta CLp^4 and δCDp4\delta CDp^4 in these zones is mainly due to the displacement of the shockfoot (or shock-feet if two shocks are not based at the trailing edge) ; (c) this numerical divergence is transferred to the adjoint components via the inverse matrix of the source terms ; (d) this does not necessarily prevent the numerical satisfaction of the adjoint lift- (resp. drag-) boundary condition at the wall (2) as the calculation of ψ2nx+ψ3ny\psi_{2}n_{x}+\psi_{3}n_{y} in this approach involves the product of δCLp4\delta CLp^{4} (resp. δCDp4\delta CDp^{4}) by (unx+vny)(un_{x}+vn_{y}).

The method of characteristics for 2D inviscid supersonic flow is a classical method for deriving ordinary differential equations and, potentially, explicit algebraic relations satisfied along two families of curves, denoted C+\mathcal{C}^{+} (left running with respect to (w.r.t.) a streamline) and C−\mathcal{C}^{-} (right running). Here, we recall the derivation of the continuous equations and study their counterparts for the (AE) equations.

For the flow, the method starting point is the Cauchy problem posed for the Euler equations: knowing the state variables along a fixed curve (L), is it possible to calculate their partial derivatives, in both space directions, for all points of (L) (this being a necessary condition for the flow calculation in the fluid domain) ? The flow variables in two neighboring points of (L), denoted here aa and bb, are linked by fluid dynamics equations and basic first order Taylor formulas.

The more general presentations deal with rotational flows and, geometrically, both planar and axisymmetrical flows [12, 13, 14, 15]. The authors derive a system of equations for the derivatives of selected primitive variables. This system is linear in the unknown partial derivatives with non linear functions of the state variables as coefficients. When its determinant is equal to zero, it cannot be solved. That is the case if bb is on the streamline of aa or, in case of a supersonic flow, if the angle of ab→=(dx,dy)\overrightarrow{ab}=(dx,dy) w.r.t. the streamline passing through aa is ±sin⁡−1(1/M)\pm\sin^{-1}(1/M) (MM being the local Mach number). The specific curves (L) where these conditions are satisfied for every points are, for all Mach numbers, the streamtraces, and, where the flow is supersonic, the so-called C+\mathcal{C}^{+} (left running curves w.r.t. the streamtraces with angle sin⁡−1(1/M)\sin^{-1}(1/M)) and the C−\mathcal{C}^{-} (right running curves the w.r.t. streamtraces with angle −sin⁡−1(1/M)-\sin^{-1}(1/M)). Along these curves, the physical existence and boundedness of the vector of unknowns allows to conclude from the nullity of the determinant in the denominator of Cramer’s formulas, to the nullity of the determinants appearing in the numerators, and this, for all the variables. The classical computational method for supersonic flow [12, 13, 16, 14, 17, 15] is supported by the corresponding differential forms valid along the C+\mathcal{C}^{+} and C−\mathcal{C}^{-} curves, and the property of constant total enthalpy and constant entropy along the streamtraces. If the flow is irrotational, homoenthalpic and homoentropic, simpler equations are derived for the velocity magnitude and the velocity angle [13, 18, 16] or the velocity potential [17] and the differential forms satisfied along the C+\mathcal{C}^{+} and C−\mathcal{C}^{-} curves may be integrated. This permits to establish the well-known equations

k−=ϕ+ν(M) is constant along a C−k+=ϕ−ν(M) is constant along a C+(4)k^{-}=\phi+\nu(M)\text{ is constant along a }\mathcal{C}^{-} \qquad k^{+}=\phi-\nu(M)\text{ is constant along a }\mathcal{C}^{+} \tag*{(4)}

in which ϕ\phi is the streamline angle, ν(M)\nu(M) the Prandtl-Meyer function,

ϕ=tan⁡−1(v/u),ν(M)=γ+1γ−1tan⁡−1(γ−1γ+1(M2−1))−tan⁡−1(M2−1),\phi=\tan^{-1}(v/u),\qquad \nu(M)=\sqrt{\frac{\gamma+1}{\gamma-1}}\tan^{-1}\left(\sqrt{\frac{\gamma-1}{\gamma+1}(M^{2}-1)}\right)-\tan^{-1}\left(\sqrt{M^{2}-1}\right),

and γ\gamma the ratio of specific heats (γ=75\gamma=\frac{7}{5} for diatomic perfect gas). Finally, let us recall that the integration of the corresponding ODEs along the trajectories results in the property of constant enthalpy and constant entropy. Besidés, Bonnet and Luneau indicate that the mechanical equations posed in a may be expressed in the Cartesian frame of reference rather than in the usual local frame derived from the local velocity [14]. Note that the 2D characteristic equations could also be calculated in the Cartesian frame and in an inexpert way, without taking advantage of the known properties of the streamtraces. Then the following the 8×88\times8 linear system relating the derivatives of the conservative variables would be solved:

[dx000dy0000dx000dy0000dx000dy0000dx000dyAB][(∂ρ/∂x)(∂ρu/∂x)(∂ρv/∂x)(∂ρE/∂x)(∂ρ/∂y)(∂ρu/∂y)(∂ρv/∂y)(∂ρE/∂y)]=[ρb−ρaρub−ρuaρvb−ρvaρEb−ρEa0.0.0.0.](5)\left[ \begin{array}{cccccccc} dx & 0 & 0 & 0 & dy & 0 & 0 & 0\\ 0 & dx & 0 & 0 & 0 & dy & 0 & 0\\ 0 & 0 & dx & 0 & 0 & 0 & dy & 0\\ 0 & 0 & 0 & dx & 0 & 0 & 0 & dy\\ A & & & & B & & & \end{array} \right] \left[ \begin{array}{c} \left(\partial\rho/\partial x\right)\\ \left(\partial\rho u/\partial x\right)\\ \left(\partial\rho v/\partial x\right)\\ \left(\partial\rho E/\partial x\right)\\ \left(\partial\rho/\partial y\right)\\ \left(\partial\rho u/\partial y\right)\\ \left(\partial\rho v/\partial y\right)\\ \left(\partial\rho E/\partial y\right) \end{array} \right] = \left[ \begin{array}{c} \rho^{b}-\rho^{a}\\ \rho u^{b}-\rho u^{a}\\ \rho v^{b}-\rho v^{a}\\ \rho E^{b}-\rho E^{a}\\ 0.\\ 0.\\ 0.\\ 0. \end{array} \right] \tag*{(5)}

The starting point of our analytical development resides in the observation that (5) and the corresponding linear system for the (AE) equations, (6), have the same determinant. From this observation, the adjoint Euler characteristics equations are established in Sec. 2. The theoretical findings are linked with former researches and illustrated by numerical computational solutions over a very fine grid in Sec. 3. Conclusions are drawn in Sec. 4.

Adjoint characteristic equations for 2D supersonic flow

The method exposed in [17] (resp. [15]) for potential (resp. general) inviscid flow has served as a guideline to our derivation for the adjoint system. For all our calculations, we assume an ideal gas law for the static pressure p=(γ−1)ρe=(γ−1)(ρE−0.5ρ∥U⃗∥2)p=(\gamma-1)\rho e=(\gamma-1)(\rho E-0.5\rho\lVert\vec{U}\rVert^2) with a constant γ\gamma.

Problem statement

References

  1. [1]Jameson, A. Aerodynamic design via control theory. Journal of Scientific Computing 3(3), 233–260 (1988).
  2. [2]Anderson, W. and Venkatakrishnan, V. Adjoint design optimization on unstructured grid using a continuous adjoint formulation. In AIAA Paper Series, Paper 1997-0643. (1997).
  3. [3]Anderson, W., and Venkatakrishnan, V. Adjoint design optimization on unstructured grid using a continuous adjoint formulation. Computers & Fluids 28, 443–480 (1999).
  4. [4]Giles, M. and Pierce, N. Adjoint equations in CFD: Duality, boundary conditions and solution behaviour. In AIAA Paper Series, Paper 97-1850. (1997).DOI
  5. [5]Giles, M. and Pierce, N. Analytic adjoint solutions for the quasi-one-dimensional Euler equations. Journal of Fluid Mechanics 426, 327–345 (2001).DOI
  6. [6]Baeza, A., Castro, C., Palacios, F., and Zuazua, E. 2d Euler shape design on non-regular flows using adjoint Rankine-Hugoniot relations. AIAA Journal 47(3), 552–562 (2009).DOI
  7. [7]Couquel, F. and Marmignon, C. and Rai, P. and Renac, F. Adjoint approximation of nonlinear hyperbolic systems with non-conservative products. In XVII International Conference on Hyperbolic Problems: Theory, Numerics, Applications, Bressan, A., Lewicka, M., Wang, D., and Zheng, Y., editors, 385–392. AIMS, (2018).
  8. [8]Lozano, C. Singular and discontinuous solutions of the adjoint Euler equations. AIAA Journal 56(11), 4437–4451 (2018).DOI
  9. [9]Peter, J., Renac, F., and Labbé, C. Analysis of finite-volume discrete adjoint fields for two-dimensional compressible Euler flows. Journal of Computational Physics 449, 110811 (2022).DOI
  10. [10]Lozano, C. Watch your adjoints! lack of mesh convergence in inviscid adjoint solutions. AIAA Journal 57(9), 3991–4006 (2019).DOI
  11. [11]Peter, J. Contributions to discrete adjoint method in aerodynamics for shape optimization and goal-oriented mesh adaptation. University of Nantes. Mémoire pour Habilitation à Diriger des Recherches, (2020).
  12. [12]Ferri, A. Application of the method of characteristics to supersonic rotational flow. Technical Report 841, NASA, (1946).
  13. [13]Shapiro, A.H. The Dynamics and Thermodynamics of Compressible Fluid Flow. The Ronald Press Company, (1954).
  14. [14]Bonnet, A. and Luneau, J. Aérodynamique. Théories de la dynamique des fluides. Cepadues, (1989).
  15. [15]Dé léry, J. Traité d’aérodynamique compressible. Volume 3. Collection mécanique des fluides. Lavoisier – Hermès Science, (2008).
  16. [16]Liepmann, H.W. and Roshko, A. Elements of Gasdynamics. Dover Publications, (1956).DOI
  17. [17]Anderson, J. Modern Compressible Flow (third edition). McGraw-Hill series in Aeronautical and Aerospace Engineering, (2003).
  18. [18]Sears, W.R. (Under the editorial direction of). General theory of high speed aerodynamics. Vol VI High speed aerodynamics and jet propulsion. Princeton University Press, (1954).
  19. [19]Venditti, D. and Darmofal, D. Grid adaptation for functional outputs: Application to two-dimensional inviscid flows. Journal of Computational Physics 176, 40–69 (2002).DOI
  20. [20]Todarello, G., Vonck, F., Bourasseau, S., Peter, J., and Désidéri, J.-A. Finite-volume goal-oriented mesh-adaptation using functional derivative with respect to nodal coordinates. Journal of Computational Physics 313, 799–819 (2016).DOI
  21. [21]Sartor, F. Unsteadiness in transonic shock-wave/boundary layer interactions: experimental investigation and global stability analysis. PhD thesis, Université Aix-Marseille, (2014).
  22. [22]Sartor, F. and Metcalf, C. and Sipp, D. Stability, Receptivity, and Sensitivity Analyses of Buffeting Transonic Flow over a Profile. AIAA Journal 53(7), 1980–1993 (2010).
  23. [23]Hirsch, Ch. Numerical Computation of Internal and External Flows: The Fundamentals of Computational Fluid Dynamics (second edition). Butterworth – Heineman. Elsevier, (2007).
  24. [24]Jameson, A., Schmidt, W., and Turkel, E. Numerical solutions of the Euler equations by finite volume methods using Runge-Kutta time-stepping schemes. In AIAA Paper Series, Paper 1981-1259. (1981).
  25. [25]Peter, J., and Dwight, R. Numerical sensitivity analysis for aerodynamic optimization: a survey of approaches. Computers and Fluids 39, 373–391 (2010).DOI
  26. [26]Peter, J., Renac, F., Dumont, A., and Méheut, M. Discrete adjoint method for shape optimization and mesh adaptation in the elsA code. status and challenges. In Proceedings of 50th 3AF Symposium on Applied Aerodynamics, Toulouse, (2015).
  27. [27]Mohammadi, B. and Pironneau, O. Applied Shape Optimization for Fluids. Oxford Univ. Press, May (2001).
  28. [28]Alauzet, F. and Pironneau, O. Continuous and discrete adjoints to the Euler equations for fluids. Int. J. Numer. Meth. Fluids 70, 135–157 (2012).
  29. [29]Lozano, C. A note on the dual consistency of the discrete adjoint quasi-one dimensional Euler equations with cell-centred and cell-vertex central discretization. Computers and Fluids 134-135, 51–60 (2016).
  30. [30]Vassberg, J. and Jameson, A. In pursuit of grid convergence for two-dimensional Euler solutions. Journal of Aircraft 47(4), 1152–1166 (2010).DOI

Paper details

vlnc:W_sdjtzrv9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents