NUMERICAL ANALYSIS AND SCIENTIFIC COMPUTING FOR PDEs AND THEIR CHALLENGING APPLICATIONS J. Haataja, R. Stenberg, J. Periaux, P. Raback and P. Neittaanmaki (Eds.) c CIMNE, Barcelona, Spain 2007 ROTATIONAL SYMMETRY, LOW DISSIPATION AND USE OF EIGEN DIRECTIONS IN WEIGHTED LSKUM S M Deshpande∗ Anil N, Arora K, K S Malagi Engineering Mechanics Unit Jawaharlal Nehru Centre for Advanced Scientific Research Bangalore - 560064, India Email: smd@jncasr.ac.in Web page: http://www.jncasr.ac.in/emu/smd/ Department of Aerospace Engineering Indian Institute of Science Bangalore - 560012 India Abstract. The KFVS (Kinetic Flux Vector Splitting) and LSKUM (Least Squares Kinetic Upwind Method) since their inception have long gone into applications in various laboratories of India. This work presents the recent advancements made in KFVS and LSKUM. The theoretical basis of KFVS has been further established by casting it in the fibre bundle frame work. This explains why KFVS works so well. The violation of rotational invariance by LSKUM, like many other multidimensional solvers, has been addressed by developing KUMARI (Kinetic Upwind Method Avec Rotational Invariance). The dissipation inherent in the first order KFVS is reduced by developing modified KFVS (mKFVS). It gives a control on the dissipation present in the scheme. The robustness is an extremely important aspect in flow problems for complex geometries. The weighted LSKUM precisely enhances the robustness of LSKUM by reducing the multidimensional least squares formulae for spatial derivatives to 1D formulae by appropriate choice of weights. Key words: KFVS, LSKUM, mKFVS, wLSKUM, KUMARI, fibre bundles 1 INTRODUCTION 1 2 The KFVS and LSKUM have long ago gone into applications in various laboratories of India. Problems involving flow past various bodies of practical interest 3 have been successfully computed . Continuing the research work to improve these numerical schemes, we have given here a brief account of recent advances made. In this work the theoretical basis of KFVS, the property of rotational invariance, low dissipation and robustness have been addressed. The theoretical basis of KFVS has been furthered by casting it in the frame work of fibre bundles. This explains why KFVS works so well and poses some interesting questions. As with many other multidimensional upwind solvers, the LSKUM violates rotational invariance though 1 Numerical Analysis and Scientific Computing for PDEs 4 the effect is seen much less in case of LSKUM compared to others . In multidimensional case the standard LSKUM uses dimension by dimension splitting which 5 results in symmetry breaking and consequently rotational invariance is lost . We have developed a rotationally invariant kinetic upwind gridfree method called KUMARI (Kinetic Upwind Method Avec Rotational invariance) to address the above mentioned issue using the connection between directional derivative, divergence and Fourier series. The first order KFVS is found to be dissipative. Though the higher order KFVS has overcome this problem, we have developed a modified KFVS (mKFVS) which achieves higher accuracy though being strictly first order accurate. The mKFVS allows the control of dissipation and gives nearly second order accurate results. The robustness is a very important aspect when one deals with flow computations past bodies of practical interest. The LSKUM which is robust, as suggested by our experience over a decade, has been made even more robust by developing weighted LSKUM (wLSKUM). The weighted LSKUM chooses the weights in the least squares evaluation derivatives in such a way that the multidimensional least squares formulae are reduced to 1D formulae. This reduces the problem of bad con3 nectivity caused code divergence . First we start with a very brief introduction to kinetic schemes and LSKUM. 2 Basic Idea of kinetic schemes 1 Kinetic schemes are based on the well known fact from the kinetic theory of gases that the suitable moments of Boltzmann equation lead to Euler or Navier Stokes equations depending upon whether the molecular velocity distribution function is a Maxwellian or Chapman-Enskog distribution. When the molecular velocity distribution function f is a Maxwellian distribution F the 1D Boltzmann equation is given by ∂f ∂f +v = 0, f = F (1) ∂t ∂x where v is the molecular velocity. The Maxwellian distribution F is given by, ρ F = I0 Df I β 2 2 exp −β(v − u) − π I0 (2) where the notations have standard meaning. The molecular velocity v, can be split into positive and negative parts using CIR splitting as, v= v + |v| v − |v| + = v+ + v− 2 2 (3) The 1D Boltzmann equation given by Eq.(1) can then be written in CIR split form as, ∂f ∂f ∂f + v+ + v− = 0, f = F (4) ∂t ∂x ∂x Now taking Ψ moments, we obtain the standard KFVS Euler equations, ∂GX + ∂GX − ∂U + + =0 ∂t ∂x ∂x (5) S M Deshpande et al./ Rotational symmetry, low dissipation and weighted LSKUM where GX ± are split fluxes given by, ± GX = Ψ, v ± F = Z R+ Z ΨvF dvdI R± T v2 where, Ψ = 1, v, I + 2 (6) By upwind differencing the spatial derivatives in Eq.(4) using finite differences, we obtain the semi discrete Boltzmann equation as, + ∂f (v f )i − (v + f )i−1 (v − f )i+1 − (v − f )i + =0 (7) + ∂t ∆xi,i−1 ∆xi+1,i Now using a forward time stepping and then taking the Ψ moments, we obtain a numerical scheme for Euler equations based on KFVS, + − GXi+ − GXi−1 GXi+1 − GXi− n+1 n Ui = Ui − ∆t (8) + ∆x ∆x The KFVS method thus involves development at two levels - Boltzmann level and Euler level, which are connected by moment relations. 2.1 Least squares Meshless Methods In the least squares based meshless methods the discrete approximation to derivative is obtained by using the least squares principle. Consider the 1D distribution of points as shown in the Fig.1. We want to evaluate Fx at P0 given Fi at each of the nodes in the connectivity set C(P0 ), which is defined as C(P0 ) = {Pi , ∀i ∈ (1, ..n) / d(Pi, P0 ) < h}, where d(Pi, P0 ) is the Euclidean distance between Pi and P0 and h is a characteristic length to be chosen by the user. Expanding the function F in Taylor series around P0 we have ∆x2i Fxx0 + O(∆x3i ) 2 where, ∆Fi = Fi − F0 , ∆xi = xi − x0 (9) Define error, E = Σ(∆Fi − Fx0 ∆xi )2 (10) ∆Fi = Fx0 ∆xi + and minimising E we get, (1) Fx0 = LS where, LS h ∂() ∂() i C " ∂F ∂x C(P0 ) # = Σ∆Fi ∆xi Σ∆x2i (11) is the notation for least squares evaluation of derivative over the connectivity C. The Fx given by Eq.(11) is first order accurate and can be easily 5 made second order accurate by defect correction technique . 2.2 Least Squares Kinetic Upwind Method (LSKUM) Least Squares Kinetic Upwind Method (LSKUM) is a gridfree kinetic upwind 2 method developed by Ghosh and Deshpande . The method uses the KFVS Euler equations and least squares evaluation of derivatives. The flux derivatives in KFVS Numerical Analysis and Scientific Computing for PDEs Figure 1: Point distribution and stencil splitting in case of 1D Euler equations are evaluated using least squares approximation which requires only a cloud of points and hence the method is gridfree. Consider the CIR split 1D Boltzmann equation given by Eq.(4). The spatial derivatives in this equation are discretised using least squares approximation and then the moments are taken to arrive at LSKUM state update formula. The Euler equations being hyperbolic, we have to evaluate the derivatives in an upwind way and accordingly split stencils are used, and these are defined by, C1+ = {Pi ∈ C(P0 )/∆xi < 0} and C1− = {Pi ∈ C(P0 )/∆xi > 0} (12) These are shown in Fig.1. Now CIR split Boltzmann equation can be upwind discretised using least squares method along with stencil splitting, " # " # ∂f ∂ + ∂ − + LS (v f ) + LS (v f ) =0 (13) ∂t ∂x ∂x C + (P0 ) C − (P0 ) 1 1 Taking the Ψ moments and using forward time stepping for the time derivative in the above equation, we obtain the LSKUM state update for 1D Euler equations as, ( U0n+1 = U0n − ∆t LS 3 " ∂ (GX + ) ∂x C1+ (P0 ) # + LS " ∂ (GX − ) ∂x C1− (P0 ) #) (14) Fibre Bundles It is very interesting to observe that KFVS can be interpreted according to the theory of fibre bundles. We have used the concepts, terminology and notation 6 of Isham . First let us introduce various spaces. Let E be space of all velocity distribution functions f (t, x1 , v1 , I) in case of 1D and f (t, x1 , x2 , v1 , v2 , I) in case of 2D. We will deal with 2D hereafter. Obviously we need to impose the constraints that every f ∈ E should have finite ψ-moments (Eq.(6)). We may further demand that f ≥ 0 (strong positivity condition) or ρ, p ≥ 0 (weak positivity condition). Note that ρ, p are unique functions of U =< ψ, f >. In the terminology of fibre bundles, E is a total space. Now introduce M =base space and the map π : E→M (15) M is the space of U =< ψ, f >. We thus have 3-tuple (E, π, M). The inverse image π −1 (U) of projection map π is the fibre over U ∈ M. Thus we have, E, π, M, ṽ = π −1 (U), B (16) S M Deshpande et al./ Rotational symmetry, low dissipation and weighted LSKUM Where E= total space, π= map, M = base space, ṽ= fibre and B= bundle. We will not worry right now in what exact sense E and M are topological spaces. For each U ∈ M, the fibre π −1 (U) is a conservative fibre in the sense that; ∀f ∈ π −1 (U) the mass, the momentum, and energy densities are the same. There is a well known 7 result that among all f ∈ π −1 (U) the Maxwellian F ∈ π −1 (U) has maximum entropy measure or equivalently minimum H, i.e., H(F ) ≤ H(f ), f ∈ π −1 (U) (17) Here H(f ) is the Boltzmann H-function which is a measure of information. Hence all the Maxwellians F ∈ E constitute a cross-section of bundle B and KFVS method for Euler equations operates in this Maxwellian cross-section of bundle B. Let us denote this cross-section by ṽ(f ), then for 2D we have the result ρβ I 2 2 F = (18) exp −β(v1 − u1 ) − β(v2 − u2 ) − I0 π I0 which gives G =< vψ, F > ∀F ∈ v(F ) (19) Hence under the map F →U (20) π : vF → G and therefore 2D Boltzmann equation maps to 2D Euler equations of gas dynamics. It is important to emphasize that for all f ∈ E but f 6= F the Euler equations are not moments of Boltzmann equation. Only for a distribution function in the Maxwellian cross-section of bundle, the Boltzmann equation maps to Euler equa1 tions. KFVS precisely uses this mapping (called moment method strategy ). There are many interesting questions about the fibre interpretation of KFVS method. Is it a principal fibre bundle, can we generate a fibre π −1 (U) using orbit of some group G? The collision dynamics occurring in Boltzmann equation gives us natural group for transforming precollision velocities vi , wi to post collision velocities vi′ , wi′ involving two parameters, b(impact parameter) and angle ǫ which plane of motion makes with some reference plane. Also there is concept of inverse collision which has as precollision velocities vi′ , wi′ and gives as post collision velocities vi , wi . In the present context, we consider these as transformations (and not collisions), i.e., vi′ , wi′ are transformed velocities and hence vi′ = vi′ (vi , wi , b, ǫ), wi′ = wi′ (vi , wi , b, ǫ) (21) Thus any two particle distribution functions of the type f (vi )f (wi ) can be transformed to f (vi′ ), f (wi′ ) and we can obtain one-particle velocity distribution function from the product f (vi′ )f (wi′ ) by integrating over wi. Because of summational invariants the resultant one-particle velocity distribution function will also belong to π −1 (U) but will have reduced value value of H−function or equivalently increased entropy. Thus it is possible to subject f (v) ∈ π −1 (U) to several collisional transformations and take it to Maxwellian F (v) ∈ π −1 (U) which has maximum entropy. This is pictorially shown in Fig.2. We end this discussion with KFVS update regarded as a fibre connection which involves movement across fibre. In KFVS update, we start with U n and get U n+1 , thus the update causes movement from fibre π −1 (U n ) to fibre π −1 (U n+1 ). We thus move from one Maxwellian to another in the Maxwellian cross-section of the bundle. Numerical Analysis and Scientific Computing for PDEs Maxwellian cross−section F f’’ Towards decreasing H Fibre −1 π (U) looping is due to entropy preserving transformation f’ M U H(F) < H(f’’) < H(f’) Figure 2: Sketch of fibres and cross section 4 Rotational invariance 5 Deshpande has studied the question of symmetry of differential operators and 1 their discrete approximations. In kinetic schemes one often deals with differential operator ∂f ∂f D (f ) = v1 + v2 (22) ∂x ∂y which arises in Boltzmann equation. In terms of the concepts of differential geometry, ∂ ∂ D = v1 + v2 (23) ∂x ∂y is a vector field and D(f ) represents rate of change of f in the direction of the field. Alternatively, we can consider 1-form df = ∂f ∂f dx + dy ∂x ∂y (24) and the vector (v1 , v2 ), and then regard D(f ) as an inner product between (v1 , v2 ) and 1-form df . The differential operator D (f ) being a scalar product is invariant under 2D rotation group O (2). Let us now consider least squares approximations 2 (1) (1) fx0 and fy0 to fx0 and fy0 respectively at some point P0 . Then the discrete differential operator is (1) (1) DD (f ) = v1 fx0 + v2 fy0 (25) 5 which as Deshpande has shown is invariant under O (2). Here, O (2) is a symmetry group of D (f ). However, when we use standard finite difference scheme to discretise fx0 and fy0 , then the resultant approximation on a 2D regular mesh is not invariant under O (2). It is however invariant under a subgroup of O (2) and this property 5 is called symmetry breaking by Deshpande , who has also shown that upwinding 2 generally breaks symmetry! For example, upwinding is enforced in LSKUM by stencil division. The upwind approximation to D (f ) used in standard LSKUM is not invariant under O(2). However it invariant under subgroup E ∈ O (2) consisting of E = {R(0), R(π/2), R(π), R(3π/2)} (26) Thus upwinding along coordinate directions based on stencil division is symmetry breaking. A question therefore arises; Can we construct an upwind approximation to D(f ) defined by Eq.(22) which is also rotationally invariant i.e., no symmetry S M Deshpande et al./ Rotational symmetry, low dissipation and weighted LSKUM breaking is allowed? In the present work we have addressed this question and have developed a rotationally invariant kinetic upwind gridfree method. The study of directional derivative reveals an interesting connection between directional derivative, divergence and Fourier series. Using this connection effectively we have developed a rotationally invariant kinetic upwind gridfree method called KUMARI. 4.1 Relation between directional derivative and Fourier series 8 It can be easily shown . that the expression for directional derivative is in fact a Fourier series with only three terms given by ∂ ~ (Q ℓ̂) = A0 + A1 cos θ + B1 sin θ + A2 cos 2θ + B2 sin 2θ (27) ∂s ∂ 1 ∂ ∂ 1 ∂ (GX) + (GY ) , A2 = (GX) − (GY ) , A0 = 2 ∂x ∂y 2 ∂x ∂y 1 ∂ ∂ B2 = (GX) + (GY ) and A1 = B1 = 0 2 ∂y ∂x ~ is the flux vector in 2D and ℓ̂ is a unit vector making an angle θ with where Q x−axis and s is the distance along ℓ̂, ~ = GX î + GY ĵ, ℓ̂ = cos θî + sin θĵ Q (28) One important observation regarding the first term of the Fourier series is in order. 1 ∂ ∂ 1 ~ ∇Q (29) A0 = (GX) + (GY ) = 2 ∂x ∂y 2 From the above equation it is clear that the term A0 contains the divergence of ~ Thus the Fourier expansion of directional derivative contains the flux vector Q. ~ information of divergence of flux vector Q. 4.2 Basic theory of KUMARI The Euler equations given by, ∂ ∂ ∂U + (GX) + (GY ) = 0 ∂t ∂x ∂y (30) can now be re-written using Eq.(29) as, ∂U + 2A0 = 0 ∂t If we have an approximation for A0 then Eq.(31) becomes an ODE, (31) dU + 2(discretised A0 ) = 0 (32) dt The above system of equations can be solved by any of the standard methods available to solve ODEs. Hence if we can find a consistent and rotationally invariant approximation to A0 , then we can devise a rotationally invariant scheme for Euler equations. For simplicity of explanation let us introduce the following notations, ~ ℓ̂ = special flux , ζ(θ) = ∂ (E) = directional derivative E=Q ∂s (33) Numerical Analysis and Scientific Computing for PDEs 4.3 Least squares Fourier fit Following the earlier discussion showing the relation between Fourier series and directional derivative, it is clear that we can represent directional derivative by Fourier series. Let us approximate ζ(θ) by Fourier series (say of three terms), ζ(θ) = A0 + A1 cos θ + B1 sin θ (34) We can approximate the above Fourier co-efficients using the well known idea of least squares fitting of functions. Suppose we have m values of ζ(θj ) for (j = 1, ..., m), where m is the number of directions which can be arbitrary. Obviously if m > 3, the above system of equations becomes overdetermined. We can then easily find A0 using the least squares principle. Further using the simple Euler time stepping yields the update formula, U0n+1 = U0n − 2∆tA0 (35) It is also possible to evaluate the Fourier co-efficients using the standard integral 8 formulae to find the co-efficients of Fourier series and numerical quadrature . 4.4 Least squares evaluation of directional derivative This section explains how to compute ζ(θ). Note that ζ(θ) is a derivative along ~ along ℓ̂. Since the Euler equations are any direction ℓ̂, of component of flux vector Q hyperbolic in nature we need to evaluate ζ(θ) in an upwind way along any direction θ. For this we need to split the special flux E into split fluxes E + and E − . The splitting is done along s. This has been achieved by developing special flux vector 8 splitting(SFVS) . Using SFVS the directional derivative becomes, ζ(θ) = ∂ ~ ∂E ∂E + ∂E − Q ℓ̂ = = + ∂s ∂s ∂s ∂s (36) To compute the derivatives of special split fluxes in the above equation we can again use the least squares approximation of derivatives, as in standard LSKUM, along with special split fluxes. The upwinding is achieved by stencil division, as in LSKUM, along each direction. The least squares formulae employed to evaluate spatial derivatives in Eq.(36) are given in [8]. It is thus clear that the split derivatives in the expression for the directional derivative, ζ(θj ) can be determined in an upwind way by using the least squares formula along each direction θj , for j = 1, . . . . . . , m, m being number of directions which can be arbitrary. 4.5 Results We have successfully solved the standard 2D shock reflection problem involving supersonic flow using KUMARI. The Fig.3(Left) shows the plot of variation of pressure along x-axis at y = 0.35 using standard LSKUM, KUMARI and exact solution. We can easily see that the KUMARI captures the shock accurately. The convergence characteristics Fig.3(centre) of KUMARI are as well better when compared to standard LSKUM. The Fig.3(right) shows the plot of pressure contours for 2D Riemann problem of 3rd standard test case of Peter Lax. We can see that slip lines, Kelvin-Helmholtz instability and other features have been captured well. S M Deshpande et al./ Rotational symmetry, low dissipation and weighted LSKUM 0 4 1 KUMARI 3.5 KUMARI LSKUM LSKUM -5 0.8 RESIDUE pressure 3 EXACT SOLUTION 2.5 0.6 -10 0.4 2 0.2 1.5 -15 1 0 1 2 3 X 0 5000 10000 N 15000 0 0 0.2 0.4 0.6 0.8 1 Figure 3: (Left)Pressure at y=0.35 (centre)residue comparison (right)Pressure contours for 2D Riemann problem 5 mKFVS The usual first order kinetic scheme is found to be more dissipative. However, 3 higher order accurate schemes have been constructed for better accuracy and for considerably reducing numerical dissipation. To control the dissipation in the first 9 order KFVS scheme, Deshpande has defined a modified way of CIR splitting, namely MCIR splitting of the molecular velocity by introducing a dissipation control function. 5.1 MCIR splitting Using Taylor series, the semi-discrete upwind scheme for the 1D Boltzmann equation, Eq.(7) gives the modified partial differential equation (mpde), ∂F ∆x ∂F +v = |v|Fxx + O ∆x2 (37) ∂t ∂x 2 In the above equation the leading term in the truncation error shows that it is first order accurate and is dissipative. The mpde Eq.(37) shows that |v| contributes to 9 the numerical dissipation. To reduce the numerical dissipation, Deshpande has introduced a dissipation control function φ as a multiplying factor for |v|. The modified way of CIR splitting of the molecular velocity, called MCIR splitting is given by v + |v|φ v − |v|φ + (38) v = v+ + v− = 2 2 Using the above defined MCIR splitting the semidiscrete upwind scheme for 1D Boltzmann equation is given by v + |v|φ Fj − Fj−1 v − |v|φ Fj+1 − Fj ∂F + + =0 (39) ∂t j 2 ∆x 2 ∆x 9 The mpde analysis of the above equation gives ∂F ∆x ∂F +v = |v|(φF )xx + O ∆x2 ∂t ∂x 2 Taking the Ψ - moments, we arrive at Z ∂G ∆x ∂U Ψ|v|(φF )xxdvdI + O ∆x2 + = ∂t ∂x 2 R+ ×R (40) (41) Numerical Analysis and Scientific Computing for PDEs In terms of φ the modified split fluxes are easily defined as Z v ± |v|φ ± Gm = ΨF dvdI 2 (42) R+ ×R Obviously, from Eq.(41) as well as Eq.(42) it follows that φ = 1 gives Gm± = G± KF V S , the usual KFVS fluxes, while φ = 0 leads to a central difference scheme. Thus, by tuning φ such that 0 < φ ≤ 1 we can control the numerical dissipation and hence order of accuracy. 5.2 Mathematical models for φ Consider the mpde obtained in the Eq.(41). When Ψ = 1, the mpde corresponding to the mass balance equation is given by " # r Z i 2 β 1 −β(v−u)2 −I/I0 ∂ ∂2 h ∆x ∂ρ ∂(ρu) |v|φρ + = 2 e dvdI = 2 νnum ρ (43) ∂t ∂x ∂x 2 π I0 ∂x R+ ×R where νnum is the numerical kinetic viscosity given by r Z ∆x β 2 νnum = |v|φe−β(v−u) dv 2 π (44) R From the above equation we observe that the absolute value of velocity, |v| is multiplied by velocity dependent exponential in the integrand on RHS of Eq.(44). For the case of φ = 1, maximum contribution to the numerical viscosity comes from particles with velocities close to u, i.e., c = v − u ≈ 0. Particles with large |c| contribute very little to νnum . If we can suitably weight the particles contributing maximum to νnum , then we can reduce the numerical viscosity in the scheme. The control function φ precisely plays this role. Obviously, φ must be a function of molecular velocity v to achieve the objective, that is, to reduce νnum . We have considered two choices, α φ = e− |v| and φ = e−α|v| (45) where α could be a mesh dependent function, which we will define later. For both the choices, φ is an exponentially decaying function. Also, in the limit α → 0 ⇒ φ = 1 and α → ∞ ⇒ φ = 0, resulting in first order accurate KFVS and central differencing schemes respectively. α When φ = e− |v| , the control function φ is very small for particles with |v| close to zero. Thus low velocity particles are not allowed to contribute significantly to νnum . For |v| → ∞, φ → 1 and contribution to numerical viscosity comes from high velocity particles whose number density is very small due to Gaussian distribution. The parameter α is a characteristic velocity scale of the control function and low and high ends of |v| are controlled by α. We can introduce non-dimensional control parameter α̃ as p (46) α = α̃/ β S M Deshpande et al./ Rotational symmetry, low dissipation and weighted LSKUM The other choice for control function is φ = e−α|v| . In this case low velocity particles, that is, particles for which α|v| ≪ 1, or |v| ≪ α1 contribute most to νnum . Obviously, particles for which α|v| ≫ 1, or |v| ≫ α1 will contribute very little to kinematic numerical viscosity. Again, α−1 is a characteristic scale of the weighting function e−α|v| . Thus, it is possible to weight the velocity space suitably for reducing the dissipation. Let us understand more about low and high velocity molecules. Consider the steady 1D Boltzmann equation with a BGK - model for the collision terms ∂f = A(F − f ) (47) ∂x Here, f is the velocity distribution function, A−1 the relaxation time scale and F the local Maxwellian velocity distribution function. Using the integrating factor eAx/v we can write the solution of Eq.(47) as Zx A A ′ (x−x ) −A 0 + (48) f (x, v) = f (x0 , v)e v F (x′ )e v (x −x0 ) dx′ , v > 0 v v x0 Similar formula can be written down for v < 0, but that is not required for making the main point in the following argument. Assuming F (x′ ) to be a constant over the interval x ≤ x′ ≤ x0 (which is ≃ ∆x), we get i h −A (x−x0 ) −A (x−x0 ) v v f (x, v) = f (x0 , v)e (49) +F 1−e It is clear from the above equation that the low velocity molecules are almost always lost (in the sense of loss term in the Boltzmann equation) while high velocity molecules are lost negligibly, that is, they travel over ∆x without any collision. Therefore, it makes sense in using a weight function in velocity space which has α − |v| is therefore consistent with the above the above property. The choice of φ = e −α|v| physical argument. The second choice φ = e leads to much simpler formulae for 9 split fluxes . 5.3 Results The mKFVS scheme has been applied to standard convergent-divergent nozzle problem. The results are compared with KFVS, and standard second order accurate MacCormack scheme. Fig.4(left) shows the pressure distribution through the nozzle. It has been observed that the mKFVS scheme is less dissipative and captures the shock more accurately when compared to KFVS and the results are in very good agreement with standard second order accurate MacCormack scheme. The mKFVS was then applied to standard transonic airfoil test case. It can be easily inferred from pressure contours Fig.5 and Cp plot Fig.4(right) that the mKFVS is less dissipative and captures the flow features more accurately. 6 wLSKUM In LSKUM for 2D, the system of equations that has to be solved to obtain the value of the derivatives is (50) A (∇F )To = b Numerical Analysis and Scientific Computing for PDEs Pressure distribution through the nozzle Cp plots of kfvs and mkfvs methods 1 1.5 kfvs mkfvs MacCormack 0.9 kfvs mkfvs 1 0.8 0.5 0.6 −Cp p/p−inf 0.7 0 0.5 0.4 −0.5 0.3 −1 0.2 0.1 0 0.5 1 1.5 2 2.5 3 −1.5 x 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 x Figure 4: (Left)Variation of Pressure along nozzle (right) Cp plot kfvs mkfvs Figure 5: Pressure contours for transonic airfoil test case (Left) KFVS (right) mKFVS where, P P P 2 ∆x ∆F F ∆x ∆x ∆y T i i xo i i i P &b= P , (∇F )o = A= P Fy o ∆yi ∆Fi ∆xi ∆yi ∆yi 2 (51) The least squares matrix A obtained above has several interesting mathematical and geometrical characteristics. First, the matrix A is purely a geometric matrix containing the differentials of coordinates of nodes in the connectivity. This matrix has to be inverted to find the derivatives at a node. This directly explains the importance of the connectivity of the node in grid-free solvers. Second the matrix A is a real symmetric matrix. We know that a real symmetric matrix has all real eigen-values and a complete set of real & distinct eigenvectors. Thus we can obtain an orthogonal basis. Third, if we rotate the coordinate frame from (x, y) to (x′ , y ′ ) such that each of the new coordinate directions is an eigenvector of A, we will obtain a diagonalized matrix A′ . It is interesting to note that in this new rotated frame, the least squares formulae for the derivatives reduce to the one dimensional formulae along each new eigendirection. Considering the two dimensional example, the least squares formula for the x-derivative in (x, y) frame is P P P P ∂F ∆y i 2 ∆xi ∆F i − ∆xi ∆y i ∆y i ∆F i (52) = Fxo = P P P ∂x o ∆xi 2 ∆y i 2 − ( ∆xi ∆y i )2 P Along the new coordinate directions, ∆x′i ∆yi′ = 0, so the above formula reduces to P P P ∂F ∆y ′ i 2 ∆x′ i ∆F i ∆x′ i ∆F i P P P = (53) = ∂x′ o ∆x′ i 2 ∆y ′ i 2 ∆x′ i 2 which is one dimensional formula along the x′ direction. S M Deshpande et al./ Rotational symmetry, low dissipation and weighted LSKUM Let us study the advantages in using 1D formula. Consider a point distribution obtained from uniform structured but highly stretched cartesian grid (∆y ≪ ∆x). The derivative at some point Po considered from such a point distribution using least squares method with the connectivity points containing standard 9 point stencil is given by Eq.(52). It is easily observed that in the least squares matrix A for such a connectivity the cross term P product P in 2the matrix A vanishes. The eigen-values of the matrix are now ∆xi 2 and ∆yi . Since ∆y ≪ ∆x the matrix A is highly ill conditioned. Use of 2D formula Eq.(52) for such a case leads nearly to 0/0 singularity because X X X ∆yi 2 ≪ ∆xi 2 , ∆xi ∆yi = 0 (54) P Even when ∆xi ∆yi does not exactly vanish, it can be vanishingly small and therefore the numerator and the denominator in Eq.(52) becomes difference between two small numbers, thus leading to loss of accuracy in the estimate of the derivative. However, use of 1D formula P P ∆xi ∆F i ∆y ∆F i Fxo = P Fy o = P i 2 (55) 2 , ∆xi ∆y i is free from such problem. Thus the same connectivity of points which was unable to give accurate value of the derivative due to the ill conditioned matrix A now gives accurate value of the derivative. The 1D formula works perfectly fine on a highly stretched cartesian mesh. So the idea is to reduce the 2D least squares formulae to 1D formulae by diagonalization of matrix A, then ill conditioning of A does not pose any problem. 6.1 LSKUM with rotation along the Eigen directions The eigen directions offer some advantages in least squares formulation. Consider the 2D split Euler equation ∂G+ ∂G− ∂G+ ∂G− ∂U y y x x + + + + =0 ∂t ∂x ∂x ∂y ∂y (56) Each of the spatial derivative in Eq.(56) when discretized by the least squares method has its unique least squares matrix A depending upon the split stencil used to calculate the derivative. As a result, it is not possible to use the one dimensional formula simultaneously for all the spatial derivatives. This can however be achieved by weighted least squares along with use of the appropriate weights such that the x and y directions become the eigen directions along which the higher dimensional least squares formulae reduce to the corresponding one dimensional formulae. The weighted and the unweighted least squares formula for the derivatives is derived in an exactly similar manner. Minimising the weighted sum of the squares of the error, n X 2 (57) E= wi ∆F i − ∆xi Fxo − ∆y i Fy o i=1 where wi is the weight assigned to each node, with respect to Fxo and Fy o as before, we get the following system of equations to be solved A(w) (∇F )To = b (w) (58) Numerical Analysis and Scientific Computing for PDEs where P P P w ∆x ∆F Fx wi (∆xi )2 wi ∆xi ∆yi T i i i P , b (w) = P , (∇F )o = A (w) = P Fy wi ∆yi ∆Fi wi ∆xi ∆yi wi (∆yi )2 (59) It is desirable to have positive weights so as to retain the LED property of the least squares formulae. It is interesting to note that an appropriate choice of the weights can even change the nature of the matrix A. So an interesting question naturally arises : Whether suitable weights can be chosen which favourably change the condition number of the least squares matrix such that the solution accuracy and robustness is improved ? The answer is affirmative. We will show later that the weights can be suitably determined such that the weighted least squares matrix A (w) is diagonal. It has been observed earlier that the diagonalization of the least squares matrix reduces the multidimensional least squares formulae of the derivatives to the one dimensional least squares formulae in the appropriate direction. It is expected that A (w) with suitable weights will help in overcoming the problems of 10 bad connectivity to a great extent . 6.2 Calculation of weights for two dimensional least squares formulae Consider x-y coordinates about a point of interest where we want to evaluate the derivatives. It will have the standard four quadrants (I, II. III & IV numbered anticlockwise). It is observed that the product of ∆x and ∆y is always positive in quadrants I and III, while it is always negative in quadrants II and IV. Whenever we are using x-y splitting, each split stencil involves two quadrants. One of the quadrants always contributes to the positive product ∆x∆y while the other quadrant always contributes to the negative product ∆x∆y. Suppose we want to find the weights for the points in the left stencil only. It comprises of quadrants II and III. Making use of the above observation, we can easily obtain the weights such that P II+III wi (∆xi ∆y i ) = 0 while ensuring positivity of the weights. Let wII be the weight assigned to the points lying in the quadrant II of the stencil while wIII be the weight assigned to the points lying in the quadrant III of the stencil. We then enforce X X wIII ∆xi ∆y i + wII ∆xi ∆y i =0 (60) III II In terms of the above quadrant wise cross products, we get P ( II ∆xi ∆y i ) wIII =− P (61) wII ( III ∆xi ∆y i ) P P As per the observations made before : ( II ∆xi ∆y i ) < 0 & ( III ∆xi ∆y i ) > 0. From the Eq.(61) above, it is seen that the ratio of the weights obtained is always positive as the product ∆x∆y is positive in quadrant III while it is negative in quadrant II. The similar procedure can be applied to find the weights for the derivatives using the points in any of the other split stencil : right, top or bottom. 6.3 Results The method has been applied to standard subsonic airfoil test case and the pressure contours have been plotted in Fig.6. We can see that the flow features are S M Deshpande et al./ Rotational symmetry, low dissipation and weighted LSKUM Figure 6: (Left)Pressure contours for subsonic airfoil test case (right) Pressure contours for multipass and singlepass results overlapped captured well. The wLSKUM was then applied to 4t h standard configuration of turbine blade 3rd test case. The Fig.6 shows the pressure contours obtained for multipassage computations and multipassage results generated by single passage computations. It clearly shows that we have been able to simulate multipassage computation with single passage computation. 7 Conclusions The recent advances made in KFVS and LSKUM have resulted in a significant improvement over the existing the algorithms. They have resulted in improvement of accuracy, robustness and Hi-fidelity representation of the governing partial differential equations. The theory of fibre bundles provided a mathematical framework to explain KFVS. The KFVS state update is now understood as a fibre connection. Several interesting questions have been raised in the fibre interpretation of KFVS. We are also able to achieve the rotational invariance using the directional derivative and special flux vector splitting (SFVS) along each direction. The mKFVS makes it possible to control the dissipation and gives nearly second order accurate results with a first order difference formula. The use of weights in weighted LSKUM (wLSKUM) to diagonalise the least squares matrix and hence reduce the multidimensional formulae to 1D formulae has resulted in a more robust solver. It has resulted in better convergence characteristics and guaranteed the LED property of 2D LSKUM. Many developments in the same direction are under way such as optimal control of dissipation, use of weights to achieve optimal bandwidth and maximising the benefits of LSKUM by combining it with optimal shape design and problems involving multiple moving boundaries. It is a great pleasure to write this paper in honour of Oliver who has been my friend for so many years. Oliver has been friend of India and has been collaborating with a large number of Indian scientists in IISc, IIT, TIFR and other universities. It can be said that he brought some aspects of French science to India and is a very good example of international collaboration. We wish him very well in all future endeavours and would love to see him in India more often. Numerical Analysis and Scientific Computing for PDEs REFERENCES [1] S. M. Deshpande, Kinetic theory based new upwind methods for inviscid compressible flows, AIAA Paper no. 86-0275, 1986. [2] A. K. Ghosh and S. M. Deshpande, Least Squares Kinetic Upwind Method for inviscid compressible flows, AIAA Paper no. 95-1735, 1995. [3] C. Praveen, Development and Application of Kinetic Meshless Methods for Euler Equations, Ph.D. Thesis, Dept. of Aerospace Engg., Indian Institute of Science, Bangalore, India, July 2004. To appear in Computers and Fluids. [4] Keshav S. Malagi, ”Rotationally invariant kinetic upwind method (KUMARI)”, MSc(Engg)thesis, Department of Aerospace engineering, Indian Institute of Science, Bangalore, India. [5] S. M. Deshpande, Meshless method, Accuracy, Symmetry breaking, Upwinding and LSKUM, Fluid mechanics report, Report No. 2003 FM 1, Department of Aerospace Engineering, Indian Institute of Science, Bangalore, India, 2003. [6] Chris J. Isham, Modern Differential Geometry for Physicists, World Scientific Publishing Co., 1989. [7] S. M. Deshpande, On the Maxwellian Distribution, Symmetric Form and Entropy Conservation for the Euler Equations, NASA TP 2583, 1986. [8] Keshav S Malagi, P S Kulkarni, S M Deshpande, KUMARI - Kinetic upwind method with rotational invariance, Proceedings of The Eleventh Asian Congress of Fluid Mechanics(11ACFM), May 2006, Kuala Lumpur, Malaysia. A37 pp. 326- 331. [9] S M Deshpande, Anil N, Omesh Reshi Accurate numerical solution of Euler equations by optimal control of dissipation. French-Indian Workshop on Numerical Simulation, Control and Design of Aeronautical and Space Applications, INRIA, Sophia Antipolis, France, Nov 29-Dec 01, 2006 [10] Arora Konark, Rajan NKS, S M Deshpande, Weighted Least Squares Kinetic Upwind Method using Eigen Vector Basis, Proceedings of 8th Annual AeSI CFD Symposium, NAL, Bangalore, India.