In the following, we will label the particles with a,b and k, where a is usually the particle of interest, b a neighbor particle, and k can be either of them. We also use the Einstein sum convention with repeated indices, which implies a sum from 1 to 3, and we usually use i,j,l and m for the summation indices. Since we are working in a non-relativistic context and do not have to distinguish between co- and contravariant vectors, it has no particular meaning whether an index is written as a subscript or a superscript; readability considerations mostly guide this.

Our starting point is a common set of Smoothed Particle Hydrodynamics (SPH) equations [24,44,27] (1)ρa=∑bmbW¯ab (2)dv→adt=−∑bmbPa+Pbρaρb∇aW¯ab (3)duadt=12∑bmbPa+Pbρaρbv→ab⋅∇aW¯ab.

In Eqs. (2) and (3), the index a at the nabla operator indicates that it is evaluated with respect to particle a (and not b). As usual, one has the choice between either solving the continuity equation directly or obtaining the density ρa through a kernel-weighted sum over neighbors, see Eq. (1), which is our choice here. The continuity equation approach has been found to be advantageous in some recent studies [22,27], but we choose here the summation approach, since it is guaranteed to deliver a strictly positive density due to the positive definite kernels W that we are using. Alternatives to this approach will be explored in the future. The size of the kernel support is determined by the smoothing length h, P denotes the gas pressure, u is the specific internal energy, and v→ab=v→a−v→b. Often, the density estimation, Eq. (1), is performed “as locally as possible” in the sense that the kernel of a particle a is evaluated with its smoothing length, ha. To be consistent with our later choices, see Section 2.4, we choose here (4)W¯ab=W(|r→a−r→b|/ha)+W(|r→a−r→b|/hb)2and∇W¯ab=∇W(|r→a−r→b|/ha)+∇W(|r→a−r→b|/hb)2.

While we have written the above equations in a commonly used way, it is worth keeping in mind that the kernel gradient can be explicitly written (for some smoothing length h) as (5)∇aWab(h)=∂W(qab)∂qab∂qab∂r→a=∂W(qab)∂qabe^abh,where qab=|r→a−r→b|/h and e^ab=(r→a−r→b)/|r→a−r→b| is the unit vector pointing from particle b to particle a. With this, Eq.(3) can also be written as (6)duadt=12∑bmbPa+Pbρaρb(v→a⋅e^ab−v→b⋅e^ab)1hab∂W∂qab=12∑bmbPa+Pbhab ρaρb∂W∂qab(v~a−v~b),where the velocities projected on the connection line are marked with a tilde.

In gas dynamics, seemingly harmless sound waves can steepen into shock waves, and for their correct treatment, dissipation is needed. So far, however, the equations are entirely nondissipative, and to handle shocks, they need to be enhanced by mechanisms that produce sufficient (but not too much) dissipation to obtain non-oscillatory solutions in shocks. As outlined in the introduction, an often used approach is artificial viscosity, which has seen many improvements in recent times, including more accurate ways to steer dissipation parameters [17–19], applying reconstruction techniques in dissipative terms [20–22]. These new developments have also been transferred to (general) relativistic hydrodynamics [12,13,45]. These new approaches have turned out to be robust and accurate and to produce very little unwanted dissipation. In our understanding, such modern artificial dissipation approaches are on par with approximate Riemann solvers.

Despite the very good performance of modern artificial viscosity schemes, we want to explore here the case where dissipation is provided by a Riemann solver [46]. The main idea is to solve (approximately) a one-dimensional Riemann problem at the midpoint between each pair of interacting particles a and b, as sketched in Fig. 2. More specifically, we replace averages of particle values by the solution of a Riemann problem (7)Pa+Pb2→Pab∗andv~a+v~b2→vab∗,where the ∗ labels the contact discontinuity state in a Riemann problem, see e.g., [46], and the ab-index refers to the solution between state a and b. As before, the tilde denotes the velocities projected onto the line connecting two particles (8)v~a=v→a⋅e^abandv~b=v→b⋅e^ab.

With the substitutions Eq. (7) we have (9)Pa+Pb=2Pab∗andv~a−v~b=2(v~a−vab∗)so that the hydrodynamics equations, now with dissipation, read (10)ρa=∑bmbW¯ab, (11)dv→adt=−2ρa∑bVbPab∗∇aW¯ab, (12)duadt=2ρa∑bVbPab∗(v→a−v→ab∗)⋅∇aW¯ab,where we have used Vb=mb/ρb and v→ab∗=vab∗e^ab.

We use Roe’s approximate Riemann solver [47] for the star state: (13)vab∗=12((v→a+v→b)⋅e^ab+Pb−PaCRL), (14)Pab∗=12(Pa+Pb+CRL(v→b−v→a)⋅e^abCRL),where the “densitized” Roe-averaged Lagrangian sound speed (the dimension is density times velocity) is (15)CRL=cs,aρaρa+cs,bρbρbρa+ρb,and cs,k is the sound speed of particle k. If, for a moment, we ignore the terms that involve the pressure and velocity differences in Eqs. (13) and (14), we have (16)vab∗≈v→a+v→b2⋅e^abandPab∗≈Pa+Pb2.

By inserting these expressions into Eqs. (11) and (12) we obviously recover the inviscid Eqs. (2) and (3), therefore the terms involving the differences in Eqs. (13) and (14) are responsible for the dissipation. One may thus try to cure potentially excessive dissipation by modifying these terms. We approach this issue here by applying the reconstructed pressure and velocity values in the dissipative terms. So, for perfectly reconstructed smooth flows, where the reconstructed values on both sides of the midpoint are the same and, therefore, the dissipative terms vanish, one effectively solves the inviscid hydrodynamics equations.

Our strategy to reduce dissipation is to use, in the dissipative terms of Eqs. (13) and (14), values of P and v→ that are reconstructed to the midpoint of each particle pair, r→abmid=0.5(r→a+r→b), see Fig. 2, left panel. So, at each interparticle midpoint, one has P/v→ values that are reconstructed once from the a-side and once from the b-side. Explicitly, the linearly reconstructed velocity components read (17)vai,rec=vai−12Ψ(∇vai,∇vbi)⋅r→ab, (18)vbi,rec=vbi+12Ψ(∇vai,∇vbi)⋅r→ab,the specific internal energy is (19)uarec=ua−12Ψ(∇ua,∇ub)⋅r→ab, (20)ubrec=ub+12Ψ(∇ua,∇ub)⋅r→ab,and the density (21)ρarec=ρa−12Ψ(∇ρa,∇ρb)⋅r→ab, (22)ρbrec=ρb+12Ψ(∇ρa,∇ρb)⋅r→ab.

With the reconstructed values of internal energy and density, we can calculate the corresponding pressure values via our equation of state, P=(γ−1)ρu.

The quantity Ψ in the above reconstruction equations is a suitable slope limiter function. Commonly used slope limiters are minmod (23)Ψmm(x,y)=12[SGN(x)+SGN(y)] MIN(|x|,|y|),the vanLeer limiter [48] (24)ΨvL(x,y)={2xyx+yif xy>00otherwise,the vanLeer monotonized Central (vanLeerMC) [48] (25)ΨvLMC(x,y)={SGN(x) MIN[|x+y|2,2|x|,2|y|]if xy>00otherwise,and the vanAlbada limiter [49] (26)ΨvA(x,y)={(x2+ϵ2)y+(y2+ϵ2)xx2+y2+2ϵ2if xy>0,0otherwise.

The latter is insensitive to the exact value of ϵ, here, we use ϵ2=10−6. In Eqs. (19) to (22), the slope limiter Ψ is to be applied to each component of the gradients.

One of the main criticisms of the standard SPH method is that it, in general, does not exactly reproduce constant or linear functions. In the standard SPH-discretization the approximation of a function f(r→) reads (27)f~(r→)=∑b Vb fb W(r→−r→b,h(r→))=∑bfb Φb(r→),where we have abbreviated mb/ρb as Vb and the smoothing length h should be thought of as a function of r→. After the second equal sign, we have abbreviated the product VbW(r→−r→b,h(r→)) as weight function Φb(r→). As an example, if all function values would be the same, fb=f0, then the function approximation should yield f0, but the standard SPH approximation finds (28)f~(r→)=f0∑bmbρbW(r→−r→b,h)≈f0,which is approximately, but not exactly, equal to the desired value, because the sum is not guaranteed to yield exactly unity. More systematically, one can expand fb (i.e., the function f at the position b) around r→, insert this into Eq. (27) to find (29)f~(r→)=∑b[f(r→)+(∇f)(r→)⋅(rb→−r→)+h.o.t.] Φb(r→)=f(r→)∑bΦb(r→)+(∇f)(r→)⋅∑b(rb→−r→)Φb(r→)+h.o.t.,where higher order terms are abbreviated as “h.o.t”. This implies, not too surprisingly, that for a good approximation, f~(r→)≈f(r→), the lowest order consistency relations (30)∑bΦb(r→)=1and∑b(r→b−r→)i Φb(r→)=0,should be fulfilled. Standard SPH usually fulfils this well in initial conditions (where particles are often placed on some type of lattice), but it does not enforce these conditions during evolution. The conditions are well fulfilled when good kernels (e.g., of the Wendland family [28]) with large neighbour numbers are used [18], but this, of course, comes at the price of expensive sums over many neighboring particles.

The consistency relations Eq. (30), however, can also be enforced by construction, e.g., in the so-called reproducing kernel method [50]. One can, for example, enhance the kernel functions with additional parameters A and Bi, (31)𝒲ab(r→ab)=Aa[1+Bai (r→ab)i]W¯ab,where W¯ab is given by Eq. (4). One can then determine the four unknown numbers A and Bi by enforcing the four discrete consistency relations so that (32)∑bVb 𝒲ab=1and∑bVb (r→ab)i 𝒲ab=0,is fulfilled at every particle position r→a. Since these kernels 𝒲ab reproduce by construction constant and linear functions exactly (i.e., to floating point accuracy), they are usually referred to as (linearly) reproducing kernels (RPKs). Reconstructing kernels for higher-order polynomials could be designed similarly, but at the price of lengthy and computationally expensive-to-evaluate expressions. We, therefore, restrict ourselves here to linear order. Since the Bi are in general non-zero, the kernels 𝒲ab are not guaranteed to be radial as standard SPH kernels, and therefore, angular momentum conservation is difficult to enforce exactly. In SPH, angular momentum conservation is usually a consequence of the interparticle forces pointing along the connecting vectors between a particle a and a particle b, e^ab, together with the kernel being anti-symmetric, ∇aWab=−∇bWab, see Section 2.4 in [8] for a detailed discussion. It is, however, still possible to write a set of equations that conserves energy and momentum (and mass if density summation is used), but not necessarily angular momentum. In practice, however, this does not seem to be a major concern; for example, the authors of [20] find with their artificial viscosity-based RPK-SPH approach in typical tests violations of exact angular momentum conservation on the sub-percent level.

The gradient of 𝒲ab can be calculated in a straight-forward way as (33)(∇a)k𝒲ab=Aa Bak W¯ab+Aa(1+Bai(r→ab)i)∇akW¯ab+(1+Bai(r→ab)i)W¯ab(∂kA)a+Aa (r→ab)i (∂kB)ai W¯ab.

Now, taking the nabla operator concerning particle b, one finds (34)(∇b)k𝒲ba=Ab Bbk W¯ba+Ab(1+Bbi(r→ba)i)∇bkW¯ba+(1+Bbi(r→ba)i)W¯ba(∂kA)b+Ab (r→ba)i (∂kB)bi W¯ba=Ab Bbk W¯ab+Ab(1+Bbi(r→ab)i)∇akW¯ab+(1−Bbi(r→ab)i)W¯ab(∂kA)b−Ab (r→ab)i (∂kB)bi W¯ab,where we have used W¯ab=W¯ba, r→ba=−r→ab and ∇bkW¯ba=−∇akW¯ab. The parameters A and Bi and their derivatives, which are needed for ∂k𝒲ab, can then be calculated by straightforward algebra. We first define discrete moments at position r→a as (35)(M0)a≡∑bVb W¯ab, (36)(M1i)a≡∑bVb (r→ab)iW¯ab, (37)(M2ij)a≡∑bVb (r→ab)i (r→ab)j W¯ab.and their derivatives read (38)(∂kM0)a≡∑bVb ∇akW¯ab, (39)(∂kM1i)a≡∑bVb[(r→ab)i(∇akW¯ab)+δkiW¯ab], (40)(∂kM2ij)a≡∑bVb [(r→ab)i (r→ab)j (∇akW¯ab)+(r→ab)i δjk W¯ab+(r→ab)j δik W¯ab].

With the moments and their derivatives at hand, one can calculate the kernel parameters (41)Aa=1(M0)a−(M2ij)a−1 (M1i)a (M1j)a, (42)Bai=−(M2ij)a−1 (M1j)a,and their somewhat lengthy but otherwise straightforward calculable derivatives (43)∂kAa=−Aa2[(∂kM0)a−2(M2ij)a−1(M1j)a(∂kM1i)a+(M2il)a−1(∂kM2lm)a(M2mj)a−1(M1j)a(M1i)a],and (44)∂kBai=−(M2ij)a−1 (∂kM1j)a+(M2il)a−1 (∂kM2lm)a(M2mj)a−1(M1j)a.

With the linearly reproducing kernels now at hand, we can approximate a function F via (45)F~(r→a)=∑bVb Fb 𝒲ab,and its derivative via (46)∂kF~(r→a)=∑bVb Fb ∂k𝒲ab.

While these expressions look very similar to standard SPH approximations, they exactly reproduce linear functions on a discrete level, which the SPH equations do not.

One desirable quantity for numerical conservation is the anti-symmetry of the kernel gradient (50)∇aWab(hab)=−∇bWab(hab),which in typical SPH equations guarantees, in a straightforward way, the numerical conservation of energy and momentum, since in the time derivative of their total values, all terms cancel exactly, as seen, for example, in Section 2.4 in [8]. The usual standard kernel gradients point in the direction of the line connecting two particles, see Eq. (5); therefore, together with the kernel gradient anti-symmetry, the exact conservation of angular momentum can also be guaranteed. Due to the presence of the vector Bi in Eq. (31) the kernels 𝒲 are no longer guaranteed to be radial, and therefore, angular momentum conservation cannot be guaranteed in the same way as in standard SPH. The standard kernel gradient as it enters Eqs. (11) and (12) can also be written as (51)∇aW¯ab=∇aWab(ha)−∇bWba(hb)2,where we have made use of the anti-symmetry of the kernel gradients. We now replace the gradients of the kernels W by the much more accurate gradients of the kernels 𝒲, see Eqs. (33) and (34), or, in other words, we replace ∇aW¯ab by (52)(∇𝒲)ab≡12[∇a𝒲ab−∇b𝒲ba],so that our final equation set reads (53)ρa=∑bmbW¯ab, (54)dv→adt=−2ρa∑bVbPab∗(∇𝒲)ab, (55)duadt=2ρa∑bVbPab∗(v→a−v→ab∗)⋅(∇𝒲)ab.

We also use the gradients (∇𝒲)ab in the reconstruction process described in Section 2.3. As mentioned before, our equation set does not manifestly conserve angular momentum as standard SPH does, but our approach produces excellent results, at least for the broad set of challenging test cases that we show below. We integrate our evolution equations forward in time utilizing a second order total variation diminishing Runge-Kutta scheme [53]. While this works very well in all tests, it is worth stating that time integration can, in principle, introduce numerical non-conservation. For detailed discussions of the role of time integration schemes in energy conservation and for compatible energy discretization schemes, we refer to the recent literature [20,54,55].

We start with a three-dimensional shock-tube-type problem. We choose the same parameters as [46] (apart from a shift of the origin): the computational domain is [−1,1]3 and the initial conditions are chosen as: (56)(ρ,v→,P)={(1.000,0,0,0,1.0)forr<0.5(0.125,0,0,0,0.1)else.

The solution exhibits a spherical shock wave, a spherical contact surface traveling in the same direction, and a spherical rarefaction wave traveling toward the origin. As initial conditions, we simply placed the 2003 particles on a cubic lattice within [−1,1]3, together with the surrounding “frozen” particles as the boundary condition.

We show in Fig. 6 the values of density, velocity and pressure for the vanAlbada limiter. Despite the initial setup on a cubic lattice, the results are practically perfectly spherically symmetric. The results for the different limiters are merely identical in this case, only on very close inspection, one finds a reminiscence of the grid structure in the case of the vanLeerMC limiter. In Fig. 7, we show our particle results in a strip around the x-axis (|y|<0.018,|z|<0.018) compared with a 4003 grid cell calculation with the Eulerian weighted average flux method [46]. Overall, we find excellent agreement. This figure can be compared to Fig. 13 of [21], where the same test was performed at the same resolution but with the MAGMA2 code using matrix inversion gradients and an artificial viscosity prescription that also uses slope-limited reconstruction. For more details on the latter method, see the section “Matrix Inversion Method I” in [21]. Although both approaches show excellent agreement with the reference solution, the RPK approach of this paper yields a sharper resolved contact discontinuity.

In a second spherical blast wave problem [46] we start from (57)(ρ,v→,P)={(1.0,0,0,0,2.0)forr<0.5(1.0,0,0,0,1.0)else.and place again 2003 particles in the same straightforward way as in the first blast wave problem.

We show the numerical solution of this test in Fig. 8. Again, deviations from sphericity are minute, and the agreement with the reference solution (Eulerian weighted average flux method with 4003 grid cells [46]), see Fig. 9, is excellent. In Fig. 10, we zoom in on the density distribution to illustrate the effect of the different slope limiters. Overall, the agreement is very good, but as expected, the minmod limiter is the most diffusive one. The vanLeerMC limiter captures best the edges of the rarefaction wave, but at the price of a small overshoot at the shock front. The results can again be compared to the “Matrix Inversion method I” (see Figs. 14 and 15 in [21]). Overall, the results are very similar, but the RPK-hydrodynamics does not show a density overshoot in the central region as in the older result.

The Sedov-Taylor explosion test, a strong initial point-like explosion expanding into a low-density environment, has an analytic self-similarity solution [56,57]. For an explosion energy E and a density of the ambient medium ρ, the blast wave propagates after a time t to the radius r(t)=β(Et2/ρ)1/5, where β depends on the adiabatic exponent of the gas (≈1.15 in 3D for the Γ=5/3 we use). In the strong explosion limit, the density jumps in the shock front by a factor of ρ2/ρ1=(Γ+1)/(Γ−1)=4, where the numerical value refers to our chosen value of Γ. Behind the shock, the density drops rapidly and finally vanishes at the center of the explosion.

To set up the test numerically, we distribute 2563 SPH particles according to a Centroidal Voronoi Tessellation (CVT) [−0.5, 0.5] × [−0.5, 0.5] × [−0.5, 0.5]. Although this produces already reasonably good initial conditions, they can be further improved by additional sweeps according to our “Artificial Pressure Method” (APM) [21]. The main idea of the APM is to push each particle to a position where it minimizes its density error. Practically, this is achieved via the following steps: i) start from a trial particle distribution, ii) measure the density at every particle via Eq. (1), iii) calculate the error in the density compared to a desired density profile, iv) from this error construct an artificial pressure and v) use this artificial pressure in an equation very similar to the hydrodynamic momentum equation to push the particles in a direction where their error decreases. After many such iteration steps, the particles end up in locations where they optimally approximate the desired density profile. Here, we need a uniform density distribution, but the method also works very well for complicated density profiles; see, for example, Fig. 3 in [21]. Even if the differences in the particle distributions are hard to see by eye, they still improve the outcome of the Sedov test. We therefore use 500 of such APM sweeps here. Once the particle distribution is settled, we assign masses so that their density is ρ=1. This is done in an iterative way where we first assign a guess value for the masses and then measure the resulting density via Eq. (1) and subsequently correct the particle masses. The iteration is stopped once the density agrees everywhere to better than 0.5% with the desired value. The energy E=1 is spread across a very small initial radius R, and is distributed entirely as internal energy; the specific internal energy u of the particles outside of R is entirely negligible (10−10 of the central u). For the initial radius R we choose twice the interaction radius of the innermost SPH particle. Boundaries play no role in this test as long as the blast does not interact with them. We therefore place “frozen” particles around the computational volume as boundary particles.

In Fig. 11, we show the density evolution (cut through xy-pane) of the case with the vanAlbada limiter. No deviation from spherical symmetry is visible, and the numerical solution for the shock agrees very well with the exact solution (the hard-to-see black line that separates the background density (cyan) from the shocked matter). The density as a function of radius is shown in Fig. 12, where each panel shows the result for one limiter. As in the tests before, and some of the later tests, the vanAlbada limiter performs best. Although here minmod performs similarly, the vanLeerMC limiter does not seem to provide enough dissipation and leads to a noisier post-shock region and an overshoot at the shock front. Again, these results can be compared against the MAGMA2 results and -taken at face value- show slightly better agreement with the exact solution. However, this may be, at least in part, because in [21] we used a Wendland kernel with 300 neighbors for the MAGMA2 result, while here we used only 220 neighbor particles.

Kelvin-Helmholtz (KH) shear instabilities occur in a broad range of environments, e.g., in terrestrial clouds, in mixing processes in novae [58], the amplification of magnetic fields in neutron star mergers [59–61] or planetary atmospheres [62], to name just a few examples. Traditional versions of SPH have been shown to struggle with weakly triggered KH instabilities [63,64], although many recent studies with more sophisticated numerical methods yielded very good results [20–22]. We focus here on a test setup in which traditional SPH has been shown to fail, even at a rather high resolution in 2D, see [64]. We follow the latter paper (similar setups were used in [20] and [22]), but we use the full 3D code and set up the “2D” test as a thin 3D slice with N×N×20 particles (referred to as "N2"), For simplicity, the particles are initially placed on a cubic lattice. Periodic boundary conditions are obtained by placing appropriate particle copies outside of the “core” volume. The test is initialized as (58)ρ(y)={ρ1−ρme(y−0.25)/Δfor  0.00≤y<0.25ρ2+ρme(0.25−y)/Δfor  0.25≤y<0.50ρ2+ρme(y−0.75)/Δfor  0.50≤y<0.75ρ1−ρme(0.75−y)/Δfor  0.75≤y<1.00where ρ1=1, ρ2=2, ρm=(ρ1−ρ2)/2 and Δ=0.025. The velocity is set up as (59)vx(y)={v1−vme(y−0.25)/Δfor  0.00≤y<0.25v2+vme(0.25−y)/Δfor  0.25≤y<0.50v2+vme(y−0.75)/Δfor  0.50≤y<0.75v1−vme(0.75−y)/Δfor  0.75≤y<1.00with v1= 0.5, v2=−0.5, vm=(v1−v2)/2 and a small velocity perturbation in y-direction is introduced as vy=0.01sin⁡(2πx/λ) with the perturbation wave length λ=0.5. In the linear regime, a Kelvin-Helmholtz instability grows in the incompressible limit on a characteristic time scale of (60)τKH=(ρ1+ρ2)λρ1ρ2|v1−v2|,with τKH≈1.06 for the chosen parameters. For our tests, we chose again a polytropic equation of state with exponent Γ=5/3.

We show in Fig. 13 the density at t=2.3 (≈2.2 τKH) for different resolutions. Even at the lowest resolution of the 1282 particles, we see healthy growth that is already very similar to that of the better-resolved cases. The evolution of the 5122 case is shown in Fig. 14. In Fig. 15, we show the density at t = 3 for the different limiters. As expected, the minmod limiter is the most dissipative one, and the vanLeerMC limiter is once again the least dissipative. Similarly to the Sedov test case, however, one may wonder whether the vanLeerMC limiter allows for enough dissipation, since the results seem somewhat noisy.

For comparison, traditional SPH implementations struggle with this only weakly triggered instability (vy=0.01), see, for example, Fig. 9 of [64], where, even at a resolution of 5122 particles, the instability either hardly grows (for the cubic spline kernel, label “Ne512”) or much too slowly (for the quintic spline kernel, label “No512”). In Fig. 16, left panel, we show the mode growth (calculated exactly as in [64]) for three different resolutions, all using the vanLeerMC limiter. Their growth rates are compared with a high-resolution reference solution (40962 cells) obtained by the PENCIL code [65]. Even our low-resolution case with 1282 particles is reasonably close to the reference solution. To quantify the agreement with the reference solution, we calculate the quantity (61)𝒟≡1N∑pN(Mpref−Mp)2,where M is the mode growth exactly calculated as in [64] for all our data points p and the superscript “ref” denotes the corresponding value of the reference solution. For the simulations shown in Fig. 16 𝒟 is 2.84×10−2 for 1282, 7.05×10−3 for 2562 and 2.65×10−3 for 5122. In the right panel of Fig. 16, we show the impact of the slope limiter in the example of the lowest-resolution runs. The 𝒟-values for the right panel (all 1282) are 4.37×10−2 for minmod, 2.70×10−2 for vanAlbada and 2.84×10−2 for vanLeerMC. In Fig. 17, we show the growth rates for the highest resolution (5122) again for the different limiters (right panel is a zoom-in of the left one). The corresponding 𝒟-values are 6.92×10−3 for minmod, 3.61×10−3 for vanAlbada and 2.65×10−3 for vanLeerMC. So the most diffusive minmod limiter grows slowest, while the growth rates of vanAlbada and vanLeerMC are very similar (vanAlbada having a small advantage at 1282 while vanLeerMC performs slightly better at 5122).

As a variant of the Kelvin-Helmholtz instability, we also perform a test in cylindrical symmetry, similar to [66]. To this end, we place particles from a radius of RBD1=0.5 to RBD2=1.5. For the region with r<1, we use (62)ρ=1ω=2P=4+2r2,and we also use (63)ρ=2ω=1P=5+r2,where ω is the angular frequency. In addition, we impose a small radial velocity perturbation on the interface given by (64)vr=v0cos⁡(10 φ) e−(r−1)22σ02,where φ is the azimuthal angle, v0=0.02 and σ0=0.1.

In our initial setup, we placed 1.95 million particles in a uniform, close-packed lattice so that the particles have the above properties. Again, we use the full 3D code to simulate a slice with 20 layers of particles in the z-direction, we do not allow for motion in the z-direction, and we enforce the above velocities within RBD1 and RBD1+0.05 and RBD2−0.05 and RBD2 as boundary conditions.

We show in Fig. 18 the density at t=2.2, again for our three limiters. There are small artefacts at the inner and outer boundaries because of our simple treatment of the boundary conditions, but in all cases, we see healthy growing Kelvin-Helmholtz instabilities with nicely winding billows. For the most diffusive limiter (minmod), we see essentially only the triggered modes growing. In contrast, in the other two cases, “parasitic” modes have also developed, which have been triggered by the granularity of our closely-packed particles shearing against each other.

The Rayleigh-Taylor instability is a standard probe of the subsonic growth of a small perturbation. In its simplest form, a density layer ρt rests on top of a layer with density ρb<ρt in a constant acceleration field, e.g., due to gravity. While the denser fluid sinks down, it develops a characteristic “mushroom-like” pattern. Simulations with traditional SPH implementations have shown only retarded growth or even complete suppression of instability [67,68].

We again adopt a quasi-2D setup and use the full 3D code for the evolution. We place the particles on a cubic lattice in the xy-domain [−0.25,0.25]×[0,1] and use 20 layers of particles in the z-direction, and also place 20 layers of particles as boundaries around this core region. The properties of particles below y=0 are “frozen” at the values of the initial conditions, for particles with ya>1, we multiply a damping factor (65)ξ=exp⁡{−(ya−10.05)2},to the temporal derivatives so that any evolution in this upper region is strongly suppressed. We apply periodic boundaries in x− direction at x=±0.25. Similar to [20] we use ρt=2, ρb=1, a constant acceleration g→=−0.5e^y and (66)ρ(y)=ρb+ρt−ρb1+exp⁡[−(y−yt)/Δ],with transition width Δ=0.025 and transition coordinate yt=0.5. We apply a small velocity perturbation to the interface (67)vy(x,y)=δvy,0[1+cos⁡(8πx)][1+cos⁡(5π(y−yt))],for y in [0.3,0.7] with an initial amplitude δvy,0=0.025, and we use a polytropic equation of state with exponent Γ=1.4. The equilibrium pressure profile is given by (68)P(y)=P0−gρ(y)[y−yt],with P0=ρt/Γ, so that the sound speed is near unity in the transition region.

We show in Fig. 19 the results for different resolutions at t=4 for the vanAlbada limiter. At all resolutions, the instability grows and reaches similar y values, but, of course, with finer detailed structures for higher resolution. In Fig. 20, we show the medium resolution case for the different limiters, again confirming the hierarchy of dissipation levels with minmod being the most and vanLeerMC being the least dissipative limiter.

Schulz-Rinne [43] suggested a set of very challenging 2D benchmark tests. The tests are constructed so that four constant states meet at one corner, and the initial values are chosen so that an elementary wave, either a shock, a rarefaction, or a contact discontinuity, appears at each interface. During subsequent evolution, complex wave patterns emerge for which exact solutions are not known. These tests are considered very challenging benchmarks for multidimensional hydrodynamic codes [43,69–71]. Such tests have rarely been shown for SPH. We are only aware of the work by [26], who show results for one such shock test in a study of Godunov SPH with approximate Riemann solvers and as benchmarks for the MAGMA2 code [21].

Here, we investigate three such configurations. We simulate, as before, a 3D slice thick enough so that the midplane is unaffected by edge effects (we use 20 particle layers in z-direction). We place particles on a cubic lattice so that 300 × 300 particles are within [xc−0.3,xc+0.3]×[yc−0.3,yc+0.3], where (xc,yc) is the contact point of the quadrants, and we use a polytropic exponent Γ=1.4 in all tests. We refer to these Schulz-Rinne type problems as SR1-SR3 and give their initial parameters for each quadrant in Table 2. These test problems correspond to configurations 3, 11, and 12 in the labeling convention of [70]. We smooth the initial conditions through ⟨A⟩a=∑b[(mb/ρb)AbWab]/[∑b(mb/ρb)Wab] so that they have a smoothness that is consistent with the evolution code.

For all three tests, we find very good agreement with the results published in the literature [43,69–71,21]. Comparing the performance of the different limiters for test SR1 (first row in Fig. 21), we see only minor differences. All cases nicely produce the shock surfaces and all show the “mushroom-like” structure near (0.2, 0.2) with only small differences. The structure around (0.3, 0.3), however, does show some differences: it looks “mushroom-like” for the least dissipative case (vanLeerMC), but not very much so for the more dissipative limiters. Also, test SR2 agrees overall well with the results in the literature, but here the “mushroom-like” structures are different: the curls show more windings for lower dissipation, that is, the smallest number for minmod and the highest for vanLeerMC. The SR3 test also agrees well with the results in the literature, here the results for the different limiters are virtually identical.