FLIP: A LOW-DISSIPATION, PARTICLE-IN-CELL METHOD FOR FLUID FLOW
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Computer Physics Communications 48 (1988) 25—38 25 North-Holland, Amsterdam FLIP: A LOW-DISSIPATION, PARTICLE-IN-CELL METhOD FOR FLUID FLOW J.U. BRACKBILL, D.B. KOTHE and H.M. RUPPEL Los A/amos National Laboratory, Los A/amos, NM 87545, USA The FLIP (Fluid-Implicit-Particle) method uses fully Lagrangian particles to eliminate convective transport, the largest source of computational diffusion in calculations of fluid flow. FLIP is an adaptation to fluids of the implicit moment method for simulating plasmas, in which particles carry everything necessary to describe the fluid. Using the particle data, Lagrangian moment equations are solved on a grid. The solutions are then used to advance the particle variables from time step to time step. An adaptive grid and implicit time differencing extend the method to multiple time and space scale flows. Aspects of FLIP’s properties are illustrated by modeling of a confined eddy, a Rayleigh-Taylor, an unstable subsonic stream, and a supersonic jet. The results demonstrate FLIP’s instability applicability to hydrodynamic stability problems where low dissipation is crucial to correct modeling. 1. Introduction mass weighted transport of momentum and en- ergy, which is usually used, is diffusive. The as- The particle-in-cell method (PlC) was once the sigmnent of grid quantities to the particles and only method able to describe highly distorted flow back causes diffusion in momentum and energy. in two or three dimensions, for which it received “Full-particle” methods like FLIP [8] and PAL considerable attention [1]. With time, its faults [7] preserve the ability of “classical” PlC to re- became evident. PlC is noisy, and it has more solve contact discontmuities, but eliminate the numerical viscosity and heat conduction than are major soujce of numerical diffusion by using a full acceptable by today’s standards [2]. Were it not Lagrangian representation of the fluid. FLIP also for its usefulness in simulating plasmas [3], PlC incorporates a variable particle size, adaptive mesh, would be mainly of historical interest, and implicit solution algorithm, to extend its ap- The problems with PlC are clearly defined. To plicability to problems with multiple time and make PlC as accurate as competing methods, its space scales. viscosity and heat, conduction must be reduced. The properties of FLIP when modeling fluid The solutions follow one of two different strate- flows in two dimensions are examined here. A gies for reducing numerical diffusion, improved comparison of FLIP with “classical” PlC, a devel- “classical” formulations [2,4] and “full-particle” opment of the correspondence between grid and PlC [5—8].The “classical” PlC method is a par- particle solutions, and an analysis of the ringing tially Lagrangian description of the fluid. To each instability are reviewed. Several examples, includ- particle is attributed a mass and a position. The ing calculations of a confined eddy, a “full-particle” PlC is a fully Lagrangian descrip- Rayleigh—Taylor instability, a subsonic stream, a tion of the fluid. To each particle is attributed all supersonic jet, and an ion beam target are given. of the properties of the fluid, including momen- tum and energy. “classical” ~icis able to resolve contact dis- 2. The particle-in-cell method continuities because of its Lagrangian representa- tion of the mass. As the particles stream from cell In PlC methods for fluid flows, the to cell during a calculation, they transport mass Navier—Stokes equations in Lagrangian form are from cell to cell without diffusion. However, the solved on a grid, where they are approximated by OOlO-4655/88/$03.50 © Elsevier Science Publishers B.V. (North-Holland Physics Publishing Division)
26 J. U. Brackbil/ et al. / FLIP: PlC methodfor fluidflow finite differences and solved. The equations are arily describe the fluid to model convection. Each mass, cycle, all of the particle data, except mass and position, is replaced by data from the grid solu- dp/dt + PV U— ~ (1) tion of the Lagrangian flow equations. The par- momentum, tide velocity in Harlow’s PlC [1], for example, is given by p dU/dt+ vp— c7(A+2~)v. U+ v• T=O, (2) ~ (6) 71=~(au1/ax1+aLyax1} V where U~is the fluid velocity at grid vertex v, and and energy, S,, is the nearest-grid-point (NGP) assignment 2+ function. The particles are then displaced using an p dI/dt +pV• U+ A(c’. U) 1.iT• T= 0, (3) interpolated velocity from the grid, where p, U, I and p are the density, velocity, dx specific internal energy and pressure, and ~uand A ~Tj7~= U0S1(x~ xv), ~ — (7) are the shear and bulk viscosities. In both “classical” and “full-particle” PlC, where S1 is a linear assignment function. The grid convection is modeled by particles moving through data are then regenerated by projection using eq. the grid. The data on the grid are regenerated (4) and similar equations for the other fluid varia- from the particle data when there is relative mo- bles. tion between the fluid and the grid. In practice, In “full-particle” PlC, the particle data de- the regeneration requires projecting the particle scribe the fluid completely from cycle to cycle, not data on to the grid points by accumulating a just during convection. Each cycle, the particle weighted sum of particle contributions at each data are updated from the grid solution, not re- grid point. The grid density, for example, is calcu- placed. In FLIP, for example, the particle acceler- lated from ation is interpolated from the grid data, ~ ~ du~ dU (4) ã=~~Si(xvxp). V (8) p This difference between “classical” and “full-par- The particle mass, ~ is shared among the cells, tide” PlC is equivalent to the difference between c, with centroids, x~,and volumes, J’~,according a partially and a fully Lagrangian description of to the value of an assignment function, S,~,,which the fluid. The partially Lagrangian description depends upon the distance between the particle given by the “classical” PlC particles results in and the cell center. The assignment function has significant diffusion of all non-Lagrangian varia- bounded support so that it is zero for distances bles. greater than some value h (usually the mesh spac- ing). It is a monotone decreasing function of its argument, and it is normalized to one for any ,~, 3. Diffusion in “classical” PlC ~ S(x~ x~)= ~S(x~ — — x~)= 1, (5) In “classical” PlC, grid variables, such as en- C V ergy and momentum, are transported from cell to where S represents any one of the often used cell by the particles as they move through the grid. nearest-grid-point, linear, or quadratic assignment To transport energy, for example, each particle is functions [22], which are described in refs. assigned a fraction of the cell energy in proportion [3,9,10,22]. to its mass before it is moved, “Classical” and “full-particle” PlC differ in whether the particle data is preserved from cycle e = m ~ E —~-s(x0 x— ) to cycle. In “classical” PlC, the particles tempor- p p M~ P C (9)
J. U. Brackbill et aL / FLIP: PlC methodforfluid flow 27 where the cell mass, M~,is equal to ~J’~ and the stationary particles cause no change in the solu- cell internal energy, E~,is equal to ~ (The tion on the grid just as in the original PlC with superscript 0 denotes data at the beginning of the NGP [1], but with linear or quadratic interpola- time step.) The particles are then displaced to xi,, tion as well. With their methods and higher order where the superscript 1 denotes data at the end of interpolation, energy and momentum transport is the time step, and a new cell energy is computed, a continuous function of particle displacement E~= e’,, s( xp — x~) and level the noise of thetooriginal comparable PlCfinite ordinary is reduced to a difference p methods. / i \ ,~ ~ In a sense, EPIC [12] is yet another way to = ~ >Jm~,Si~x~, x~’)SI~x~xe). — — (10) improve “classical” PlC. In EPIC the diffusion, C C ~° which can be calculated from eq. (10), is canceled Even when the particles are not displaced, the by an anti-diffusion step (cf. Eastwood, these pro- new cell energy will not equal the old [2]. Consider ceedings). the simple example of evenly spaced, equal mass particles on an evenly spaced grid, x,~.1 = x, + Lix, in one dimension. 0yields,Fourier transforming eq. (10) 4. “Full particle” PlC with E1 — x~, 2 0 = x~ The success of plasma simulation, which is a k— Sk Ek. ~ ~ “full particle” PlC, has influenced Marder (GAP) With the commonly used spline interpolation [5], Tajima et al. [6], Morse et al. (PAL) [7], and functions [22], for example, the Fourier transform Brackbil and Ruppel (FLIP) [8] to try similar is given by [3], methods for fluid flows. The advantage of the “full-particle” representation is the elimination of sin( k i~X/2) “~ 1 numerical viscosity and thermal conduction. By S~(k)= k i~x/2 (12) making momentum and energy Lagrangian varia- bles, diffusion of these variables is eliminated just (Nearest-grid-point (NGP) interpolation cone- as mass diffusion is eliminated in “classical” PlC sponds to n equal to 0, linear interpolation to n by making mass a Lagrangian variable. equal to 1, and quadratic interpolation to n equal “Full particle” PlC obviously costs more, be- to 2.) For k equal to zero, S~(k)is equal to one cause for each particle a position, velocity, mass and energy is conserved. For all k greater than and energy must be stored. (The extra cost is not zero, S~ ( k) is positive but less than one. Conse- burdensome now, because increased computer quently, interpolating energy from the grid to the memory and hardware-indirect-addressing are now particles and back is diffusive. In fact, using higher available.) More importantly, since each particle order splines actually increases the diffusion. has its own velocity, interpenetration across Harlow avoids the zeroth order error in “classi- material interfaces or multistreaming can occur cal” PlC by using NGP interpolation [1]. With just as in the smoothed particle hydrodynamics NGP, only those particles residing in cell c con- (SPH) method [9]. tribute to the energy in cell c, and there is no To reduce multistreaming, Tajima et al. [6] spatial averaging. However, energy is transported calculate the particle displacement from the par- discontinuously as particles cross cell boundaries, tide velocity, but introduce a Krook collison op- contributing to noise and diffusion in the results. erator to diminish the relative velocity between Nishiguchi and Yabe [2], and Clark [4] avoid neighboring particles. This affects the results simi- the zeroth order error in “classical” PlC in a larly to a viscosity. To eliminate multistreaming, different way. They use higher order interpolation PAL and FLIP calculate the particle displace- only for the changes in momentum and energy ment, as in “classical” PlC, by interpolating from due to transport. Their formulations require that the grid velocity. This eliminates multistreaming
28 J. U. Brackbill eta!. / FLIP: PlC methodforfluidflow completely. (PAL introduces some dissipation to backward Euler time differencing, even angular suppress the ringing instability [10], which FLIP momentum is conserved by the difference equa- does not [8].) tions [11].The details are given in ref. [8], to which some improvements have since been made. Im- plicit equations are formulated similarly to the 5. The FLIP algorithm method of Casulli and Greenspan [13] using ICCG (incomplete Cholesky decomposition, conjugate The FLIP (Fluid-Implicit Particle) algorithm gradient method), with substantial improvements uses Lagrangian particles to represent the fluid, in speed and reliability. and a rezonable Lagrangian grid of quadrilateral zones to calculate the interactions among the par- 5.2. Consistent particle and grid solutions tides. Variables are assigned to the grid using the standard von Neumann, Richtmyer staggering (see FLIP is formulated to given consistency be- Crowley, these proceedings). In its mixture of a tween the solution of the Lagrangian equations of particle representation of the fluid and a grid for motion on the grid, and the equations of motion the solution of implicit moment equations, it is of the particles. Consistency is achieved, in part, similar to the implicit moment method in plasma because the normalization of the assignment func- simulation [23]. tion given by eq. (5) is dependent on the grid placement. Furthermore, the particle equations of 5.1. The Lagrangian calculation motion, eqs. (7) and (8), are consistent with the Lagrangian solution on the grid, because the assign- Using the interpolation functions discussed ment function, S. is a Lagrangian invariant, above, the specific internal energy, vertex mass dS dt 0 — 16 and center of mass velocity at the grid points are — given by, To achieve this, S is defined in terms of natural coordinates, (~,~) that are calculated by mapping = ~e~S2(x~ — x~)/p~J’~, (13) each of FLIP’s quadrilateral cells on to a unit square. With this choice, the grid defines the par- MV = >m~S1(x~ x0), — (14) tide size. As the grid and particles move, the value p of the assignment function at a point moving with / i ~ the fluid is constant. U=J~muS1I~x—x)/M. p p p v V ~15) At each vertex, the natural coordinates assume integer values, (i, f) which are obviously constant Linear interpolation is used for vertex-centered as the mesh moves. Elsewhere, the mapping be- variables, such as position, velocity and mass, to tween physical and natural coordinates is given by be consistent with the quadrilateral cells of the bilinear interpolation, which can be written, computation grid. Quadratic interpolation is used — for cell-centered variables, such as density and Xp — ~ — 1, 7J —J energy, to give smoother solutions and to suppress the ringing instability, which is discussed in sec- — S ~ — ~ S 1— J 17 tion 6. data forBy theevaluating eqs.Lagrangian solution of (4) and (13)—(15), equations the of — 1~ ~1 i’~ motion on the grid are assembled. The natural coordinates of a particle, whose physi- The solution of the Lagrangian equations of cal coordinates are xi,, are defined by inverting motion on the grid, eqs. (1)—(3), can be approxi- this mapping. Because S is a Lagrangian in- mated in many ways. In FLIP, the equations are variant, the particle’s natural coordinates are con- formulated using a finite volume method, which stant during the Lagrangian phase, and the par- conserves mass, momentum and energy. With tide displacement, eq. (7), and acceleration, eq.
J. U. Brackbill etat / FLIP: PlC methodforfluidflow 29 (8), are calculated by interpolating from the grid variation among the particles is poorly repre- values. The correspondence between particle and sented by a linear interpolation function. In this grid solutions through the Lagrangian phase is sense, it is a measure of how accurately the grid perfect for linear functions of the dependent vari- resolves the flow gradients. It is this term that is ables. described as a kinetic pressure in ref. [8]. There it From eq. (5) and eqs. (13)—(17), one can show was noted that it is large where large changes in that the total mass, momentum, angular momen- the flow direction or velocity occur. turn and internal energy for the particles and the The variance of the particle velocity about the grid are equal, mean, the first term on the right side of eq. (23), — does not change because of accelerations calcu- — ~MV, (18) lated on the grid, which have no effect on the P V relative particle velocity, = ~MVUV, (19) dv du dU p V aEV’!~v =0 ~‘25 dt dt ‘—‘dt V5~ ‘ x ui,) = ~ x lx,), (20) V P V Solving the equations of motion on the grid affects = ~ ~ (21) only those length scales that the grid resolves. The p C sub-grid scale information that the particles carry is unaffected, and can be retrieved by refining the However, the kinetic energy is not the same for zomng. the particles and the gnd. If one defines the . . . As is required for nonlinear stabihty, the FLIP particle velocity relative to the mean velocity on . . . . the grt‘d by formulation velocity at a is vertex dissipative. during aWhere time step the ischange the in ~ VP = ui,, — ~ IJVS,,J,, S~ = S(x~ xe), — (22) change in the particle kinetic energy is always V different from the change in the grid kinetic en- ergy by the amount one can show that the difference between particle and grid kinetic energy is given by = — ~ T,~~ (26) ~[~mpup2_~MpV2I VV’ . P V This error is negative defimte, particle by particle, and order ~at2. When the loss in kinetic energy is = ~ {m~v~ — ~ u. i~p~u~} , (23) added to the particle energy for each particle, V,V’ energy is conserved exactly while marntaimng positive particle internal energy. where the transfer rnatnx is defined by, = ~ ~ — (24) 5.3. Adaptive zoning The transfer matrix is symmetric and diagonally The final step in the computation cycle requires dominant and, therefore, has positive eigenvalues defining the grid for the next cycle. During this [8]. Thus, both terms on the right side of eq. (24) step, the particles are stationary in physical space, are positive definite and the difference in the but their natural coordinates must be recomputed. kinetic energy of the particles and the grid can The grid may be completely redefined from one have either sign. However, if one estimates the cycle to the next. It is only required that the grid relative size of the two terms, the second term is map on to a logical rectangle. In FLIP, the num- the result of the averaging one does to calculate ber of grid points is fixed, but the grid points may the grid velocities and is never larger. The first be moved about as needed. Presently, one may term will be relatively large when the velocity choose to use an Eulerian, fixed grid, a Lagrangian
30 J. U Brackbill et at / FLIP: PlC methodfor fluidflow grid (for a Lagrangian calculation, the particles To minimize I, one must solve Euler equations. In are superfluous), or an adaptive grid. The adaptive the solution, the weight function, w, controls the grid is generated by solving a variational problem, mesh spacing. Where w is larger, the grid points in which one minimizes a functional. The func- are closer together. If w is an appropriate function tional may be complex allowing one to control of the data, the grid will adapt to provide in- grid smoothness, orthogonality, spacing and orien- creased resolution where it is needed by moving tation [24,25]; or it may be simple, giving control grid points. For example, if w is proportional to only over spacing, such as the functional, the current density in a calculation of magnetohy- F drodynamic flow, the grid will cluster to resolve I = f v 2 ~“E+ W ~ 21 dv. (27) steep gradients in the magnetic field intensity [27]. Adaptive gridding can be useful in PlC calcula- 4 , . , - . . Fig. 1. The adaptive grid plot is superimposed on the particles representing a confined eddy. The eddy, which is rotating counterclockwise, advances one-half rotation period each frame. By the lower-right frame, it has rotated twice. A Kelvin—Helmholtz instability causes secondary eddies to form, which move and coalesce. In these, the higher kinetic pressure causes the grid points to cluster.
J.U. Brackbill et al. / FLIP: PlC methodfor fluidflow 31 tions. For example, when w is proportional to the more particles per zone, and the adaptive mesh number of particles per zone, where the particles cannot give the large gains in accuracy observed are sparse the zones will be larger. This extends with finite difference methods [24]. the range of densities that can be represented by Adaptive gridding has been useful in resolving making the particle size larger in regions of low material interfaces and in responding to changes density. When w is proportional to the kinetic in scale that may occur when shocks develop. pressure, eq. (23), the resolution is automatically Examples of adaptively zoned FLIP calculations increased where the grid does not resolve the flow are given in refs. [8,14]. gradients. The results of this prescription are shown in fig. 1, where the particles in a confined, low-speed eddy are shown superimposed on the 6. The ringing instability grid at intervals of one-half eddy rotation. The calculation is performed on a 24 x 24 zone grid With four to sixteen particles per cell in a with thirty-six particles per zone. This eddy prob- typical PlC calculation, there are many more de- lem, which is described in more detail in ref. [11], grees of freedom for the particles than there are is subject to a Kelvin—Helmholtz instability. The for the grid. Thus, the solution of the particle instability causes secondary eddies to form. equations of motion is underdetermined by the The resolution required to resolve the flow solution of finite difference equations on the grid. changes in time, and, consequently, the grid The many different particle solutions which look changes from frame to frame. The grid is initially the same when projected on to the grid are called clustered in the boundary layer between the eddy aliases. The aliases interact nonlinearly to drive and the background fluid. Later, the grid is clus- the well-known finite grid instability in plasma tered in the secondary eddies, where the gradients simulation [3] and the ringing instability in PlC in the velocity become steep and the kinetic pres- fluid calculations [10]. sure becomes relatively large. (The kinetic pres- While the assignment function for the vertex sure is less than one per cent of the fluid pressure.) centered quantities, such as position and velocity, The corresponding results on an Eulerian grid is determined by the shape of the computation with the same number of zones are different, but cells, the choice of assignment function for the the results with the same number of particles on a cell-centered quantities, such as density and en- mesh with four times as many cells (48 x 48 zones) ergy, is unconstrained. Thus, one may use higher differ only in detail from those in fig. 1. The order assignment functions to reduce the growth instability growth rate and resulting large scale rate for the ringing instability [10]. Quadratic in- structures are similar. The principal difference is terpolation is currently used in FLIP for cell- that the cost of the 48 x 48 zone calculation is centered variables. There is a disadvantage, how- twice that for the adaptively zoned calculation. ever, because the ability to resolve gradients is (The time step for the calculation is the same, as diminished using higher order interpolation [9]. allowed by the implicit formulation, but the 48 X Dispersion analysis of the ringing instability in 48 zone calculation costs twice as much with four PlC for small time steps, uniform grid spacing, times as many zones, since the grid solutions ~x and uniform flow, U0, yields the dispersion require half the total computation time.) In three relation, dimensions, where an adaptive grid would require 2 a tenth as many zones as a fixed grid calculation, 1 — Si(kq)S2(kq)k~a =0, ~ (28) the savings would be even greater. There are constraints on adaptivity in pi~ ~ — kqUo]2 — calculations. In low-speed-flow problems, the where a is the sound speed, ~ the frequency, k number of particles per zone must be sufficient to the wave number, and kq the alias wave number, give smoothly varying pressures and densities, even in the smallest zones. As a result, one must use kq = k + q(2.rr/i~x). (29)
32 J. U Brackbill et at / FLIP: PlC methodforfluidflow The imaginary frequency, labelled gamma in fig. yet for FLIP, but large time steps are observed to 2, corresponds to exponential growth for Mach be very effective in suppressing the instability if numbers, M equal to U0/a, less than 0.4. The other time step constraints allow them. minimum growth time is about ten oscillation periods (With nearest grid point interpolation re- placing S1 and S2 in eq. (28), the maximum 7. Test problems unstable M is one, and the minimum growth time little more than one oscillation period.) When the total energy is a constant of the With a new method like FLIP, confidence is motion, the ringing instability is bounded nonlin- accumulated over time by applying it to a wide early in PlC fluid calculations to relatively low range of problems. From the results, one begins to amplitudes. The only evidence of the instability is understand its properties, its strengths and its a disordering of the ordinarily regular patterns of weaknesses, in ways that complement analysis. particles, with a consequent spatial modulation of FLIP has been applied to many problems, some of the density or energy on the grid. In fig. 3, the which are: (The reference number before the slash particles representing the dense fluid in a calcula- is a reference for FLIP results, the reference after tion of the Rayleigh—Taylor instability [14] are the slash for a description of the experiment, or plotted. With M equal to 0.001, the particles other computed results.) clump into patterns that are characteristic of the 1. Interaction of a shock and a thin foil [8] ringing instability. With quadratic assignment and 2. High speed flow over a step [8/15] implicit differencing, the effect of the instability is 3. Interaction of a shock in air and a helium usually limited to unimportant effects like the one bubble [10/16] shown. Including the effect of a very large time 4. Plane shock (Sod’s problem) [14/17] step in the dispersion analysis has not been done 5. Noh’s problem [14/18] 6. Rayleigh—Taylor instability [14/19] 7. Imploding sphere and ellipse (unpubl.) 8. Confined eddy [il/Monaghan, these pro- ceedings] 9. Astrophysical jet [unpubl./20] One can make several observations about the performance of FLIP from the results of these tests. FlIP gives the correct jumps in mass, momentum and energy across the shock, but the jumps are several cells thick and not as sharp as _____ can be for method, achieved example with [15].the Because piecewise FLIP parabolic does not have numerical diffusion, it is important to choose the correct artificial viscosity. The grid Reynolds / ~ number, Reg = I u ~x/X, should be equal to one / in the shock front to give the minimum shock N u thickness The real without strength overshoot. of FLIP is its ability Both to model hydrodynamically unstable interfaces. the ~ no conservation of angular momentum and the ab- sence of dissipation appear to be essential compo- Fig. 2. The ringing instability in FLIP occurs for flow Mach nents of this capability. An illustration of the way numbers less than 0.4. A normalized growth rate of 0.1 corre- sponds to a growth time of 10 oscillations at the normal FLIP treats shear layers is given by a model frequency. problem.
J. U. Brackbill eta!. / FLIP: PlC methodforfluidflow 33 (A) (B) (C) Fig. 3. Particles are plotted depicting the dense fluid in a Rayleigh_Taylor instability. In (c), the effect of the ringing instability is apparent. ___ ___ ~~~~1 ~ (A) (B) ~ fl P (C) T~1i~ (D) Fig. 4. A subsonic stream is modeled. The velocity is depicted in (a), the initial particle placement in the background fluid in (b), the same particles plus injected particles at t = 5 s/a on a 10 x 30 zone mesh in (c), and on a 20 X 60 zone mesh in (d).
34 J. U Brackbil! et at / FLIP: PlC methodforfluid flow 7.1. Subsonic stream be seen in the relative displacement of the par- ticles, which reflects the velocity profile. (Because As shown in fig. 4a, a stream of fluid enters an the inflow velocity is constant across the aperture aperture at the bottom of the domain, flows be- at the bottom boundary and varying linearly across tween two flat plates, and exits at the top. The the stream just inside the boundary, the beam is stream and the background fluid through which it perturbed.) When the mesh spacing is halved, fig. flows have the same density and pressure. The 4d, the boundary layer thickness is halved as well, distance between the plates is five times the stream and the increased gradients in the velocity drive thickness, s, and the length of the stream is 15 s. the stream unstable. The Mach number for the flow is 1/Viö~,and the The subsequent development of the instability artificial viscosity is zero. The particles in the is displayed in fig. 5. With 36 particles per cell and upper portion of the domain at the initial time are a 10 x 30 mesh, the stream is gently meandering shown in fig. 4b. The same particles, as well as at t = 10 s/a in fig. 5a. At the same time with a those that have entered the bottom meanwhile, are 20 x 60 mesh and 9 particles per cell so that the shown at t = 5 s/a in fig. 4c. The dynamics of total number of particles is the same, fig. Sb, the these particles is calculated on a mesh with 10 X 30 stream has become turbulent. Evidently, the insta- zones, L~x= ~ y = s/ 2, and 36 particles per zone. bility can manifest itself only when the gradients With this resolution, there are just two zones of the flow are resolved by the mesh. Doubling the across the stream. Because the velocity varies un- grid spacing again (40 x 120 zones) while keeping early in space, the boundary layer includes the the number of particles per zone constant at 9 entire stream and one zone on either side as can yields the results shown in fig. 5c. The larger r HI 4 hI ... I jiil)1D)~ .?‘~ I’ I Hi... ~ 1~~~ ‘ ~ I ~ ~ ‘~1 ‘ 4. ____ (A) (B) (C) Fig. 5. Background and injected particles as in fig. 4 are plotted at t 10 s/a for a lOx 30 zone mesh, (a), a 20 X 60 zone mesh, (b), and a 40 x 120 zone mesh, (c). (a) and (b) have the same number of particles. (b) and (c) have 9 particles per cell.
J. U Brackbi!l et a!. / FLIP: PlC methodfor fluidflow 35 features of the flow are similar to those in fig. Sb, background density. The jet speed is 30 times the but more rotation and mixing has Occurred. Be- ambient sound speed, a. cause there is no viscosity there is no physical In fig. 6, the injected particles are plotted at length scale, and the results on the scale of the intervals of 2R/a. The interface is very unstable, grid will not converge as the mesh is refined. On and produces eddies on all scales resolved by the longer scales, the results should converge, mesh, as is confirmed by velocity vector plots, not shown. The size of the eddies increases with dis- tance from the head of the jet, as does the width 7.2. Supersonic jet of the turbulent flow region. The flow of the jet is unsteady. This is especially evident in figs. 6c and The unstable stream is a gentle breeze corn- d, where disruptions in the jet itself can be seen pared with the supersonic jet shown in fig. 6. (The The large density discontinuity between the jet jet shown is one of many studied by Winider and and the background gas presents some computa- collaborators [20] in modeling astrophysical jets.) tional difficulties. For example, it is difficult to The jet is injected through an aperture in the have neither too little nor too much viscosity in bottom boundary and allowed to flow freely mixed cells. To deal with this problem, the layer through the right and top boundaries. The left of mixed cells is treated as an internal boundary. boundary is an axis of symmetry. The domain has In mixed cells, where particles of two or more dimensions 6R by 16R, where R is the radius of “colors” reside, the particle contributions are given the jet, and is resolved by a 30 X 80 zone mesh, only to surrounding cells that are mixed, or con- with 5 zones in the jet. tam particles of the same color. This procedure The jet is in pressure equilibrium with the yields, typically, a band of mixed cells only one background gas, but with only one percent of the cell thick separating two materials. (A) (B) (C) (D) Fig. 6. Particles in a supersonic jet are plotted at intervals of 2 R /a. The particles are injected at the left-bottom. The left boundary is an axis of symmetry.
36 J. U. Brackbi// et a!. / FLIP: PlC method for fluid f/ow I •1 I Fig. 7. Results from a calculation of the response of a target to a heavy ion beam are plotted. Green, blue and red particles represent lead, lithium and deuterium. respectively. The target is shown at 10. 17, 20. 22, 26 and 29 ns in frames (a) through (1).
J. U. Brackbill et at / FLIP: PlC methodfor fluidflow 37 7.3. Ion-beam driven hydrodynamics 8. Current research FLIP can be applied to more complex prob- Space does not permit more than a brief discus- lems as well as to the simple fluid flow problems sion of the adaptive zoning capability of FLIP. described above. For example, one of us (DBK) is Adaptive zoning extends the dynamic range of presently using FLIP to study ion-beam-driven FLIP, not only in spatial length scales but also in targets [14]. In these, bismuth ions that have been density variation. It also improves the resolution accelerated to 10 GeV deposit 4.25 Mi of energy of shocks and other discoiitinuities. Current re- in 30 ns deep within a target. There, the ion search emphasizes understanding the strategies one energy is converted to hydrodynamic motion that should use with FLIP to adapt the grid. compresses and heats a portion of the target FLIP is presently being extended to multi- material to the densities and temperatures re- material flow with slip interfaces and material quired for fusion. properties, such as strength, and to magnetohy- To model such targets, which are very far from drodynamics. Many of the ideas in FLIP are being thermodynamic equilibrium, the physical model in used in developing a new, adaptively zoned, plasma FLIP must be extended beyond the Navier—Stokes simulation code in collaboration with D.W. Fors- equations. Some extensions require cell-based lund. A version of FLIP for three-dimensional calculations, and are simple to add. For example, flows is an attractive, and apparently feasible, to model material properties realistically one uses project for the future. complex equations of state like SESAME [21]. One also must calculate the equilibration of tern- References peratures between electrons and ions, and the ionization state of the material. Other extensions [1] F.H. Harlow, Meth. Comput. Phys. 3 (1963) 319. require calculating gradients, like ray-tracing to [2] A. Nishiguchi and T. Yabe, J. Comput. Phys. 52 (1983) model ion beam propagation, and energy trans- 390. port equations to model electron thermal conduc- [3) C.K. Birdsall and A.B. Langdon, Plasma Physics via Com- puter Simulation (McGraw-Hill, New York, 1985). tion and convection. Luckily, the underlying grid- [4] R.A. Clark, Los Alamos National Laboratory Report No. based data structure in FLIP resembles that of . LA-UR 79-1947, 1979 (unpublished). many complex physics codes, making it easy to [5] B.M. Marder, Math. of Comput. 29 (1975) 473. borrow and steal the needed physical models. [6] T. Tajima, J.N. Lebouef and J.M. Dawson, J. Comput. The results of a calculation of the dynamics of Phys. 38 (1980) 237. [7) R.L. McCrory, R.L. Morse and K.A. Taggart, Nucl. Sci. a planar, HIBALL (Heavy Ion Beam And Liquid Eng. 64(1977)163. Lithium) target are shown in fig. 7. The colors [8] J.U. Brackbill and H.M. Ruppel, J. Comput. Phys. 65 label the target materials, with green for lead, blue (1986) 314. for lithium, and red for deuterium. The ions, [9] JJ. Monaghan, Comput. Phys. Rep. 3 (1985) 71. which enter the problem from above, are focused [10] J.U. Brackbill, The Ringing Instability in Particle-in-Cell Calculations of Low-Speed Flow, J. Comput. Phys. (to be at infinity but deposit their energy in the lithium. publ.). The fluence of ions varies sinusoidally by five [11] J.U. Brackbill, Comput. Phys. Commun. 47 (1987) 1. percent across the target, with several wavelengths [12] J.W. Eastwood, Comput. Phys. Commun. 44 (1987) 73. included in the domain of the calculation. Where [13) V. Casulli and D. Greenspan, Intern. J. Numer. Meth. the intensity is highest, the lithium is pushed [14) Fluids 4 (1984) D.B. Kothe, 1001. Hydrodynamic Stability and Symmetry of aside. There, the beam penetrates more deeply and Ion-Beam-Driven Planar Targets, Purdue University the- causes a highly corrugated shock to form. The sis (1987) unpubl. highly distorted flow that results is easily modeled [15] P. Woodward and P. Colella, J. Comput. Phys. 54 (1984) by the particles. 115. [16] J.-F. Haas and B. Sturtevant, Interaction of Weak Shock Waves with Cylindrical and Spherical Gas Inhomogenei- ties, Graduate Aeronautiedl Laboratory, California In- stitute of Technology, 1986 (unpubL).
38 J. U. Brackbill et at / FLIP: PlC methodforfluidflow [17] GA. Sod, J. Comput. Phys. 27 (1978) 1. [23] R.J. Mason, in: Multiple Time Scales, eds. J.U. Brackbill [18] W.M. Noh, J. Comput. Phys. 72 (1987) 78. and B.I. Cohen (Academic Press, Orlando, 1985) pp. [19] BJ. Daly, Phys. Fluids 10 (1967) 297. 233—270. W.P. Crowley, An Empirical Theory for Large Amplitude J.U. Brackbill, ibid. pp. 271—3 10. Rayleigh—Taylor Instability, UCRL-72650, Lawrence [24] J.U. Brackbill and J.S. Saltzman, J. Comput. Phys. 46 Livermore National Laboratory (1970). (1982) 342. [20] M.L. Norman, L. Smarr, K-H.A. Winkler and M.D. Smith, [25] A.E. Giannakopoulos and AJ. Engel, Directional Control Astron. Astrophys. 113 (1982) 285. in Grid Generation, J. Comput. Phys., to be published. [21] K.F. Holian, T-4 Handbook of Material Properties Data [26] J.U. Brackbill, Solution Adaptive Grids for Time-Depen- Bases, LA-1060-MS, Los Alamos National Laboratory, dent Problems in Two and Three Dimensions, J. Comput. Los Alamos, NM (1984). Phys., submitted. [221 R.W. Hockney and J.W. Eastwood, Computer Simulation [27] R.D. Milroy and J.U. Brackbill, Phys. Fluids 25 (1982) Using Particles (McGraw-Hill, New York, 1981). 775.
You can also read