12

CHRISTOPHER K.W. TAM

(1997), Zhuang et al. (1998, 2002), Zingg (2000) and solution is damped in time. The time rate of others for use in CAA. Each of these authors offered damping for wave component with wave number a is Im[w (a )] . examples to demonstrate the numerical damping of his/her upwind scheme could, indeed, damp out A simple way to reduce numerical damping due spurious waves to produce an acceptable solution. to temporal discretization is to use a smaller Dt . To However, strictly speaking, the spatial derivative see why reducing time step size would reduce terms of the Euler equations are wave propagation numerical damping, let us assume that a spatial terms. They are not damping terms. To use the discretization scheme is chosen so that the problem discretized form of the derivative terms to perform would be solved accurately by using N mesh points two functions, namely, to support wave propagation per wavelength. In other words, if l is the shortest and to damp out spurious short waves at the same wavelength the scheme is designed to resolve, then l /Dx > N . This yields, in terms of wave number, time, might be imposing too many constraints on the 2p / N > a largest Dx . Now w is related to a by the stencil coefficients. This would make it less likely that the stencil design is optimal. A more natural dispersion relation (20a), thus strategy is to use a central difference scheme and 2p Dt c > wlargest Dt . (37) incorporate needed numerical damping by the N Dx addition of artificial selective damping terms. By choosing Dt /Dx small, wlargest Dt would be For central difference schemes, a is real for real confined to a range that is closer to zero. For most a . Such schemes do not have intrinsic damping temporal marching schemes, including the Rungearising from spatial discretization. However, a Kutta method, the DRP scheme, the w Dt versus computation scheme must have the capability to wDt relation is such that Im(w Dt) and hence eliminate spurious short waves that could be Im(wDt) becomes smaller when w Dt is closer to generated at surfaces of discontinuities, zero. It follows that the root w (a ) of (18) would be nonlinearities and computation boundaries. These nearly real when Dt is reduced. This means that the spurious short waves not only can contaminate the temporal damping rate Im(w (a )) would be computed solution, they can, in some cases, cause reduced. numerical instability. To deal with the problem of spurious short waves, Tam, Webb & Dong (1993b) designed a family of artificial selective damping 2.4. Aliasing stencils by means of the wave number analysis. These damping terms, when added to the The fundamental wave number range of a finite discretized equations, will selectively damp the high difference scheme is -p aDx < p (see Fig. 2). Wave wave number (short waves) component of the numbers that fall outside this range are undernumerical solution but practically leave the low resolved. Their wavelength is less than 2 mesh wave number (long waves) component undamped. spacings. They are aliased back inside the An alternative to artificial selective damping is to fundamental range. To determine the relationship perform filtering operation after the solution has between the original wave number and the aliased been advanced by one or more time steps. Lele wave number, let us take the initial condition of (1) (1992) proposed a compact filter. Other filters can be to be a Gaussian with a concentration of wave found in the work of Kennedy & Carpentter (1994), number around a 0 ; i.e., Visbal & Gaitonde (2001) and others. x 2 - ( ln2 ) +ia 0 x b Temporal discretization also introduces numerical . (38) t = 0 F( x ) = e damping. The damping rate can be determined The Fourier transform of (38) is quantitatively from the dispersion relation. Because (a -a 0 ) b 2 w (w ) is complex, the root w = w(a ) of - ( ln2 ) b (a ) = 2ln2 . e (39) w (w ) = ca (a ) has an imaginary part even when 1 t=0 F 2(p ln 2) 2 a (a ) is real for real a . For the convective wave equation, the exact solution of the fully discretized If b , the half-width of the Gaussian, is large, then system is given by (19). The integrals can be (39) confirms that there is an essential concentration of wave number around a = a 0 . evaluated by the Residue Theorem. This leads to the replacement of w by w = w (a ) , the root of Let Dx be the mesh size used in a computation as w (w ) = ca (a ) . Since w(a ) is now complex, the shown in Fig. 1. l is the spatial index so that