WorksPhysics and Astronomy
High-Order ADER-DG Hydrodynamics with ExaHyPE: Implementation, Validation, and Astrophysical Benchmarking
Abstract
We describe a high-order ADER-DG solver for the compressible Euler equations within the ExaHyPE framework. The implementation combines a high-order ADER-DG polynomial representation, a local space-time DG predictor, adaptive mesh refinement, and an a posteriori subcell finite-volume limiter. We test the code on a deliberately mixed set of one- and two-dimensional problems: a strong-shock Sod-type problem, the Shu-Osher shock-entropy interaction, the Woodward-Colella blast wave, a contact-driven vortex sheet, and a shock-interface interaction. The one-dimensional cases recover the expected Euler wave patterns and show clear order-dependent gains in smooth and oscillatory regions. The two-dimensional cases probe a different part of the method, namely contact preservation, shear-driven roll-up, baroclinic vorticity deposition, and Richtmyer-Meshkov-type growth. In these tests the high-order update gives the expected resolution away from discontinuities, whereas the subcell limiter keeps the calculation stable near shocks and steep interfaces. The resulting code provides a reproducible ExaHyPE implementation for idealised inviscid, non-relativistic flows in which shocks, contacts, and multidimensional interfaces are the dominant features.
equations, Shock-capturing methods, ExaHyPE, Astrophysical fluid dynamics, Riemann problem, Numerical simulation
Introduction
Compressible-flow simulations are difficult because smooth waves, shocks, and contact discontinuities can appear in the same calculation. This is especially relevant in gas dynamics and astrophysical hydrodynamics, where strong compressions and interacting interfaces are part of the flow rather than exceptional events. Mathematically, these problems are governed by hyperbolic systems of conservation laws, whose solutions may develop discontinuities even from smooth initial data [1, 2]. The numerical difficulty is therefore not only to obtain high formal order in smooth regions. The method must also remain conservative and non-oscillatory when shocks or contacts are formed. This tension has been present since the work of Godunov [3]: low-order schemes are robust but often too diffusive, while high-order schemes require additional mechanisms to control oscillations without erasing the structures one wants to resolve.
High-order shock-capturing methods address this balance directly. The ADER methodology of Toro and Titarev [4] provided a one-step framework for high-order accuracy in both space and time through the solution of Generalised Riemann Problems. Later extensions connected this idea with multidimensional nonlinear systems and with the Discontinuous Galerkin framework [5, 6, 7, 8]. For the present work, the important point is the modern ADER-DG form: a high-order DG polynomial represents the cell-local solution, a local space-time DG predictor evolves that polynomial inside each cell, and a conservative corrector communicates neighbouring cells through numerical fluxes.
The framework has also become more robust in shock-dominated regimes. The a posteriori subcell finite-volume limiter of Dumbser et al. [9] allows the high-order DG candidate to be tested after it is computed and replaced only where it becomes inadmissible. Space-time adaptive formulations with dynamic adaptive mesh refinement [10] add a second practical ingredient: resolution can be concentrated near shocks, contacts, and emerging small-scale structures. These developments have supported applications in relativistic magnetohydrodynamics, compressible Navier–Stokes and resistive magnetohydrodynamics, general-relativistic continuum mechanics, and the Einstein–Euler system [11, 12, 13, 14]. In this context, the ExaHyPE engine provides a computational environment for high-order hyperbolic solvers with parallelism and dynamic adaptivity [15]. Many published ExaHyPE applications have focused on geophysical wave propagation and relativistic astrophysical fluid dynamics [15, 16].
The non-relativistic Euler equations have received less attention in this particular setting. This absence matters for two reasons. First, the Euler system remains the natural starting point for compressible hydrodynamics and for controlled tests of shock-capturing methods. Second, the numerical ingredients that make ADER-DG attractive in more complex models—high-order polynomial representation, local space-time prediction, adaptive refinement, and subcell limiting—can be examined more cleanly before additional physics is introduced. A systematic Euler implementation in ExaHyPE is therefore not just a simplified case; it is a necessary step for understanding how these ingredients behave together. To the best of our knowledge, the published literature does not yet provide a reproducible implementation in this environment that combines a high-order ADER-DG polynomial representation, a space-time DG predictor, adaptive mesh refinement, and an a posteriori subcell limiter with a validation suite for one- and two-dimensional Euler benchmarks.
In this work, we present such a solver for the compressible Euler equations and assess its behaviour through benchmark problems of increasing complexity. In one spatial dimension, we consider the Sod shock tube [17], the Shu–Osher entropy-wave interaction [18], and the Woodward–Colella blast-wave test [19]. In two dimensions, we study interacting contact discontinuities and a shock–interface configuration that produces multiscale structures associated with Richtmyer–Meshkov and Kelvin–Helmholtz dynamics. With this set of tests, we do not only verify the implementation. We also examine how the polynomial representation, space-time prediction, limiting, and adaptivity behave when smooth waves, sharp discontinuities, and multidimensional effects coexist. Several of these configurations are also relevant to astrophysical hydrodynamics, where non-relativistic shock interactions remain an important modelling regime.
All simulation codes and datasets produced in this study are made publicly available to support reproducibility and future extensions. The paper is organised as follows. § 2 introduces the Euler equations and their characteristic structure. § 3 describes the ADER-DG method and its limiting strategy, and § 4 summarises the implementation in ExaHyPE. § 5 and § 6 present the one- and two-dimensional results. § 7 discusses the main findings, and § 8 closes the paper.
Governing Equations
The Compressible Euler Equations
We model the flow with the inviscid compressible Euler equations in conservative form. This is the form needed once shocks appear, because the solution must then be interpreted weakly and the numerical method must preserve conservation across discontinuities. The equations read [1, 2]
where is the vector of conserved variables and is the flux tensor, with in spatial dimensions. In one dimension,
while in two dimensions we use
Here is the density, are the velocity components, is the total energy per unit volume, and is the thermal pressure. The physically admissible states are those for which density and pressure remain positive,
This set is important later because the limiter is designed precisely to reject candidate updates that leave .
Equation of State and Admissible Thermodynamic States
We close the system with the ideal-gas equation of state [2, 20],
where is the specific internal energy and is the adiabatic index. The total energy density is
so the pressure recovered from conserved variables is
This is the quantity checked in the positivity criterion of the limiter. The corresponding sound speed is
which sets the acoustic propagation scale. Throughout this work we use , except in benchmark problems where another value is specified.
Hyperbolicity and Characteristic Wave Structure
For the characteristic analysis, the one-dimensional Euler equations are written in quasi-linear form as
In the admissible state space, this system is strictly hyperbolic [1, 2]. Its eigenvalues are
with whenever . The corresponding right eigenvectors form
where is the specific total enthalpy. With , the flux Jacobian is diagonalised as
This characteristic basis also fixes the local wave decomposition used by the numerical fluxes and admissibility checks. Locally, the characteristic variables
separate the solution into wave families that propagate with speeds .
This decomposition also tells us what the numerical method must capture. The acoustic fields associated with and are genuinely nonlinear, whereas the middle field associated with is linearly degenerate [21]. The outer families generate shocks or rarefactions; the middle family corresponds to a contact discontinuity. In physical terms, acoustic waves carry pressure information, while the contact wave transports material and thermodynamic jumps at the fluid speed. This distinction is the reason why the benchmark suite must include both shocks and contact-dominated configurations.
Once shocks form, the solution is no longer classical. For the present purposes, two facts are enough. Conservative formulations remain meaningful after discontinuities appear, and physically relevant shocks must satisfy an entropy admissibility criterion. For a -shock of speed , the Lax condition is
which states that characteristics of the -th family enter the shock from both sides [1, 21]. This is the continuous background behind the shock-capturing and admissibility checks used later.
In two dimensions, the same wave structure is recovered along a chosen normal direction. The quasi-linear form is
with and . For a unit normal , the projected Jacobian has eigenvalues
The system is therefore hyperbolic in every spatial direction. The middle eigenvalue is double because tangential momentum adds one material degree of freedom, but the eigenspace remains complete. This normal-direction structure is the one used in interface fluxes and in the CFL estimate of the numerical method.
The Riemann Problem and its Generalisation
The numerical fluxes are built around the same local wave structure. At a cell interface, the basic model is the Riemann problem,
where two constant states meet at a single discontinuity. In one dimension its solution consists of two acoustic waves and one contact wave: the acoustic waves may be shocks or rarefactions, while the contact carries density and entropy jumps at continuous pressure and normal velocity [2]. This is the local pattern behind the shock-tube and interaction problems used later.
This local problem also fixes the numerical point of view. Godunov’s method obtains a conservative upwind flux from it [3]. ADER keeps the same interface-centred idea, but replaces piecewise constant states with local polynomial data and evolves them in time. This Generalised Riemann Problem [22] is what lets ADER schemes reach high order in space and time without multi-stage time integration [4]. In the next section, this local view is combined with the high-order ADER-DG polynomial representation, the space-time predictor, and the conservative corrector.
Numerical Method
The numerical tests require a method that remains high order in smooth regions while staying robust at shocks and contact discontinuities. We therefore use the limiting ADER-DG formulation available in ExaHyPE: a high-order DG polynomial represents the solution inside each cell, a local space–time predictor evolves that polynomial over one time step, and a conservative corrector couples neighbouring cells through numerical fluxes. If the high-order candidate fails admissibility checks, an a posteriori subcell finite-volume limiter recomputes only the troubled cells.
High-Order DG Polynomial Representation
At time , the numerical solution in each cell is represented by a polynomial of degree . In reference coordinates , we write
where are local nodal basis functions and are the cell-local degrees of freedom. The same polynomial representation is used for each conserved variable of the Euler system. Increasing raises the formal order of the smooth-cell update, while the limiting strategy described below protects the calculation when the candidate polynomial becomes inadmissible near discontinuities.
Local Space-Time DG Predictor
The cell-local DG polynomial is then evolved locally over . The result is a space-time polynomial computed without neighbour communication. This local predictor replaces the analytic Cauchy–Kowalewski expansion of early ADER schemes [4] and is better suited to nonlinear systems [8].
With reference coordinates
the predictor is expanded as
with summation over repeated indices. Projecting , where , and integrating the time derivative by parts gives
where is the DG polynomial at . The required matrices are
The predictor degrees of freedom satisfy
with . It is solved by Picard iteration,
For the CFL-limited steps used here, only a few iterations are needed [8]. The predictor then provides the time-dependent interface traces at and for the conservative update.
ADER-DG Corrector Step
The corrector is where neighbouring cells communicate. From the left and right predictor traces, and , the cell average is advanced in one conservative step:
Here is the Rusanov, or local Lax–Friedrichs, flux, and the time integral is evaluated with Gauss–Legendre quadrature. The result is a high-order candidate: it is accepted in smooth regions but checked before being committed.
The time step follows the DG CFL restriction
where , is the smallest local cell size, and accounts for the polynomial resolution inside a cell. Unless a benchmark states otherwise, .
A Posteriori Subcell Finite-Volume Limiter
Near strong discontinuities, the high-order candidate may contain negative density, negative pressure, or non-finite values. We use the a posteriori strategy of Dumbser et al. [9]: compute the high-order candidate, test it, and recompute only the cells that fail. This keeps the limiter local and preserves the unlimited scheme where it is already admissible. The workflow is shown in Figure 1.
Figure 1. Workflow of the ADER-DG update with a a posteriori subcell finite-volume limiting. The scheme first constructs a high-order candidate, checks its admissibility, and recomputes only troubled cells on a finite-volume subgrid.
Each candidate is checked against three criteria:
Physical positivity: and , with pressure recovered from (7);
Entropy admissibility: the local entropy measure is compared with neighbourhood values at time ;
Numerical regularity: the candidate must not contain not-a-number or infinite values.
A cell that passes keeps the high-order candidate. A cell that fails is marked as troubled and recomputed on equispaced subcells [9, 10]. The previous cell average is projected to subcell averages, evolved with a second-order TVD finite-volume method using the Rusanov flux and a minmod slope limiter, and projected back to obtain . This limited state replaces the candidate only in the troubled cell. The choice lets the subcell update use the same time step as the DG scheme.
The limiter is deliberately conservative: it leaves admissible high-order updates untouched, but provides a robust fallback near shocks, contacts, and strong compressions. The troubled-cell set also acts as a diagnostic for regions requiring additional numerical control.
Adaptive Mesh Refinement
For adaptive runs, we use the dynamic mesh refinement available in ExaHyPE [15], following the space-time adaptive formulation of Zanotti et al. [10]. Refinement is driven by local variation in the DG solution, and coarsening is applied where that variation falls below the prescribed tolerance. The ADER-DG update and subcell limiter are applied independently in each AMR patch, with the local time step constrained by (26).
Fluxes across coarse–fine interfaces are corrected conservatively, so refinement does not compromise global conservation. AMR therefore concentrates resolution near shocks and contacts while keeping smooth regions inexpensive.
Implementation in ExaHyPE
The method is implemented in ExaHyPE, an open-source C++ framework for parallel, dynamically adaptive solvers for hyperbolic PDEs [15]. The framework provides pure ADER-DG, pure finite-volume, and LimitingADERDG solvers. We use LimitingADERDG because it matches the a posteriori subcell strategy described in § 3.4. Although OpenMP and MPI are available, the benchmarks reported here run on a single shared-memory node.
Solver Configuration
The Euler solver is specified through six C++ kernel functions. The routine flux evaluates the flux tensor in equations (2)–(3), with pressure recovered from (7). The routine eigenvalues returns in each coordinate direction; this is the speed used in the CFL condition (26). The routine eigenvectors provides and from (11) for the characteristic decomposition used by the Euler kernels. Finally, boundaryValues sets the boundary conditions described in § 4.2. Together, these routines define the physical part of the solver.
The limiter is controlled by the two remaining routines. mapDiscreteMaximumPrincipleObservables declares and as the monitored quantities. The routine isPhysicallyAdmissible implements the checks of § 3.4: it rejects states with , , a non-finite variable, or an entropy measure below the neighbourhood minimum at . A false return triggers subcell recomputation for that cell.
The numerical parameters common to all simulations are listed in Table 1. Polynomial degree and AMR levels are benchmark-specific and are stated in § 5 and § 6.
Table 1. Numerical parameters used unless stated otherwise. Polynomial degree and AMR levels are benchmark-specific; see § 5 and § 6.
Initial and Boundary Conditions
Initial conditions are set through adjustPointSolution, which is called at for each mesh degree of freedom. The benchmark states are given in § 5 and § 6.
Boundary conditions use one of two strategies. One-dimensional tests use transmissive outflow boundaries at both ends, implemented by copying the adjacent interior state into ghost cells so that waves can leave the domain without reflection. Two-dimensional tests use periodic conditions transverse to the shock propagation direction and transmissive conditions along the propagation direction. Both choices are implemented through boundaryValues and selected in the ExaHyPE configuration file, keeping benchmark-specific data separate from the solver kernels.
One-Dimensional Benchmark Tests
The sequence of tests is intentional. We begin with the Sod shock tube, where an exact solution exercises all three wave families. We then move to Shu–Osher, where a shock interacts with fine-scale entropy structure, and finally to the Woodward–Colella blast wave, where strong shocks reflect and collide inside a confined domain. All simulations use the parameters of Table 1. Benchmark-specific values are stated below.
Sod Shock Tube
The Sod shock tube is a Riemann problem with a known exact solution. Two uniform states meet at a diaphragm placed at . When the diaphragm is removed, the solution develops a left-going rarefaction fan, a contact discontinuity, and a right-going shock — one representative of each wave family from § 2.3. The initial conditions are in Table 2.
| State | |||
| Left | 1.0 | 0.0 | 1000.0 |
| Right | 1.0 | 0.0 | 0.01 |
Table 2. Initial conditions for the Sod shock tube. Domain , final time .
We discretise the domain with maximum mesh size and two AMR levels, giving an effective resolution of . We run four schemes at polynomial orders corresponding to third- through ninth-order accuracy. Transmissive outflow conditions are imposed at both ends.
This is a strong-shock Sod-type variant rather than the classical textbook Sod test. The pressure ratio is chosen deliberately to stress the high-order update and the subcell limiter under a compressive contrast closer to idealised blast-wave and supernova-remnant conditions.
The profiles in Figure 2 compare density, velocity, and pressure for all four orders against the exact Riemann solution at . All orders place the rarefaction, contact, and shock in the expected sequence. The rarefaction fan between and becomes sharper as the order increases, while shock and contact positions remain accurate across the set.
Figure 2. Sod shock tube profiles at for density, velocity, and pressure. ADER-DG results at orders 3, 5, 7, and 9 are compared against the exact Riemann solution.
The same comparison exposes the local trade-off near discontinuities. Around the contact at , the seventh-order solution is sharp and nearly free of diffusion; the ninth-order solution, by contrast, develops localised dispersive ripples. These Gibbs-type oscillations persist because the candidate DG polynomial remains physically admissible. The limiter intervenes only when admissibility is threatened, replacing the affected cell with a monotone finite-volume update rather than filtering every oscillation.
In smooth regions, the density error decreases consistently with polynomial order. The convergence behaviour is in agreement with the theoretical rate of for ADER-DG schemes [8]. A systematic convergence study with and errors across multiple mesh resolutions will be added in a later preprint update. Here, the Sod test is used more narrowly: it checks wave-pattern fidelity and the behaviour of the limiter under an extreme pressure jump.
The extreme pressure ratio of in this configuration mimics, in a controlled one-dimensional setting, the strong compressive fronts generated during the free-expansion phase of supernova remnants.
Shu–Osher Problem
The Shu–Osher problem extends the Riemann setting by placing a sinusoidal density perturbation ahead of a moving shock [18]. The shock propagates into this modulated medium and generates a complex post-shock oscillatory pattern. This exercises spatial resolution in a way the Sod test cannot: both a sharp front and fine-scale smooth structures must be resolved simultaneously. The initial conditions are in Table 3.
| State | |||
| Left | 1.515695 | 0.523346 | 1.805 |
| Right | 0.0 | 1.0 |
Table 3. Initial conditions for the Shu–Osher problem. Domain , discontinuity at , final time .
We run orders with AMR up to one refinement level. No exact analytical solution exists, so we use a seventh-order simulation on a 5000-cell uniform grid as the reference.
The density profiles in Figure 3 make the order dependence clear at . Third order damps the post-shock oscillations. Fifth order recovers their general shape but still underestimates peak amplitudes. Seventh order follows the reference most closely throughout the oscillatory region, although a small phase shift remains visible at all orders. It accumulates because even minor differences in numerical shock speed displace the whole post-shock pattern over five time units. The shift does not reflect inaccuracy in amplitude or frequency resolution.
Figure 3. Shu–Osher density profiles at over the full domain. ADER-DG results at orders 3, 5, and 7 are compared against a high-resolution reference solution.
The detailed post-shock structure points to the same conclusion: seventh order captures the amplitude and spacing of the oscillations, but the whole pattern sits slightly ahead of the reference. For this comparison, a phase-insensitive diagnostic is more informative than a pointwise norm.
Phase-insensitive error analysis via Shannon entropy
Standard and norms are not suitable here. A small phase shift produces large norm values even when amplitude and frequency are accurately captured. We compare the solutions through their amplitude distributions rather than their pointwise values.
For the entropy analysis we restrict the comparison to the post-shock oscillatory region, excluding the undisturbed upstream state and the shock jump itself. The restricted interval is where phase errors are most visible, but the main question is whether the numerical method reproduces the resolved amplitude distribution. For each scheme we build a normalised histogram of density values over the same spatial interval and using the same fixed binning as the reference solution. The resulting probabilities define an empirical amplitude distribution rather than a pointwise error measure. Tests with nearby bin counts give the same ordering of schemes, so the interpretation depends on the relative broadening and entropy of the distributions, not on a particular choice of bin width.
We compute the amplitude histogram of the post-shock density field for each order and for the reference. The histogram quantifies how well each scheme reproduces the statistical spread of the oscillatory signal, independent of phase. We then measure the Shannon entropy of each distribution [23],
where is the normalised frequency of the -th bin. Higher entropy corresponds to a broader and more structured amplitude distribution, hence to a larger fraction of the shock–entropy imprint being resolved. Because only the amplitude distribution enters, spatial phase is discarded by construction.
Figure 4. Phase-insensitive Shu–Osher error analysis based on density-amplitude distributions. The histogram panel compares numerical orders against the reference, while the entropy panel summarises the recovered structural diversity as Shannon entropy.
Figure 4 summarises this comparison. The third-order histogram is narrow and modal, a signature of excessive diffusion. As the order increases, the histograms broaden toward the reference distribution. The corresponding Shannon entropies, listed in Table 4, rise monotonically with order. Most of the gain occurs between orders 3 and 5; the smaller increment from orders 5 to 7 suggests diminishing returns once dispersive effects begin to compete with additional resolution. The reference entropy is the upper bound a perfect numerical solution would reach at this resolution.
| Scheme | (bits) |
| ADER-DG (order 3) | 4.21 |
| ADER-DG (order 5) | 4.67 |
| ADER-DG (order 7) | 4.84 |
| Reference | 4.98 |
Table 4. Shannon entropy of the post-shock density amplitude distribution for the Shu–Osher problem. Higher values indicate greater structural fidelity.
In this setting, Shannon entropy acts as a phase-insensitive quality metric for ADER-DG schemes in the shock–entropy interaction regime. It complements norm-based comparisons when phase drift is part of the numerical response rather than a simple loss of amplitude accuracy.
The Shu–Osher configuration models the interaction of supernova-driven shocks with preexisting density inhomogeneities in the interstellar medium, where entropy perturbations modulate the post-shock flow structure.
Woodward–Colella Blast Wave
The Woodward–Colella problem is the most demanding of the three [19]. Two strong shocks propagate inward from high-pressure boundaries, collide near the centre, reflect, and produce a transient multi-wave structure with interacting shocks, contacts, and rarefactions. The pressure ratio is . The initial conditions are in Table 5.
| Region | |
| 1000.0 | |
| 0.01 | |
| 100.0 |
Table 5. Initial conditions for the Woodward–Colella blast wave. Domain , final time . Density and velocity everywhere.
No analytical solution exists. We construct a reference solution with a seventh-order ADER-DG scheme at . That simulation required more than ninety hours of wall-clock time on a single node. The cost illustrates why matching such a reference by brute-force refinement is unattractive, especially for lower-order schemes. Production simulations use with two AMR levels.
The reference solution at is shown in Figure 5. The dominant features are two reflected shock fronts near and , contact discontinuities near –, and outward rarefaction fans near both boundaries. These structures are the benchmark for the production simulations.
Figure 5. High-resolution reference solution for the Woodward–Colella blast wave at , showing density, velocity, and pressure together with the principal wave structures.
The production runs in Figure 6 compare orders . Third order captures the global wave layout but smears shocks and contact surfaces. Fifth order places the primary shocks accurately and resolves most contacts. Seventh order follows the reference across density, velocity, and pressure, including the secondary contacts and the transient central region. The subcell limiter is active across a large fraction of the domain throughout the interaction phase. This behaviour is consistent with the problem setup: extreme pressure gradients trigger the admissibility check repeatedly, and the limiter responds locally without degrading the high-order solution elsewhere.
Figure 6. Woodward–Colella blast wave profiles at . ADER-DG results at orders 3, 5, and 7 are compared against the high-resolution reference solution for density, velocity, and pressure.
The blast wave dynamics here mirror, in one spatial dimension, the shock focusing and vorticity deposition that occur in the interior evolution of supernova remnants and stellar wind-blown cavities.
| Benchmark | Numerical challenge | Main finding |
| Sod | Rarefaction, contact, and shock at pressure ratio | All orders recover the correct wave pattern; higher order reduces smooth-region diffusion, while the highest order exposes local dispersive oscillations near discontinuities. |
| Shu–Osher | Shock interaction with entropy waves and phase-sensitive oscillations | Seventh order best reproduces the oscillatory amplitude structure; Shannon entropy provides a phase-insensitive measure of recovered fine-scale content. |
| Woodward–Colella | Strong shock interactions, reflections, contacts, and limiter activation | Seventh order remains closest to the high-resolution reference while preserving robustness in the most compressive regime tested here. |
Table 6. Summary of the one-dimensional benchmark results.
Taken together, the one-dimensional tests isolate the regime that motivates this validation: shocks and contacts coexist with entropy structure, large pressure jumps, and repeated limiter activation. In that regime, high-order ADER-DG with subcell limiting reproduces the standard Riemann wave pattern, retains post-shock oscillatory content, and remains stable in the Woodward–Colella blast wave. These results make the method a credible candidate for astrophysical flows in which the relevant structure is not always captured by pointwise errors alone.
Two-Dimensional Benchmark Tests
The one-dimensional tests in § 5 exercise wave interactions along a single axis. In two space dimensions, additional effects enter: waves can propagate obliquely, interfaces can interact geometrically, and hydrodynamic instabilities can grow from genuinely multidimensional gradients. The benchmarks in this section target these effects in two complementary limits. The first contains only contact discontinuities and tangential shear. The second adds a strong oblique shock that strikes a density interface and drives multiscale instability growth. Together, the tests extend the validation beyond colinear wave dynamics. Both calculations use polynomial order — a representative fifth-order configuration — with AMR and the subcell limiter active throughout.
Contact-Driven Vortex Sheet Dynamics
The first problem is designed to isolate contact discontinuities and tangential shear, with no shock forcing. The domain is divided into four quadrants at and , each with its own velocity state. Density and pressure are constant inside each quadrant. The density jumps define the contact surfaces, while tangential velocity jumps across those contacts provide the initial shear. Table 7 gives the corresponding states.
| Quadrant | ||||
| Top-left (, ) | 1 | 0.75 | ||
| Top-right (, ) | 2 | 0.75 | 0.5 | |
| Bottom-right (, ) | 1 | 0.5 | ||
| Bottom-left (, ) | 3 |
Table 7. Initial conditions for the vortex sheet test. Domain , interfaces at . Pressure is uniform. Four values of are studied: 1.0, 0.75, 0.5, 0.25.
The shear dynamics are described by the two-dimensional vorticity field,
and the baroclinic contribution to its evolution is
The right-hand side of (29) is the baroclinic torque. In this setup the initial pressure is spatially uniform, so at and there is no initial baroclinic source. Vorticity is instead introduced by the imposed tangential jumps at the quadrant interfaces. The subsequent motion is then a nonlinear reorganisation of these vortex sheets rather than the result of impulsive baroclinic deposition.
Four simulations are run with , while the resolution is kept fixed at with two AMR levels. Decreasing raises the Mach number of the tangential flow and moves the problem progressively farther from the nearly incompressible limit.
Figure 7 compares the four pressure values at the same final time. In each case, the oblique shear layers wind into alternating high- and low-density streaks as the flow is rotated and stretched. The value of controls both the strength and the internal structure of this pattern. For , the motion remains close to incompressible and the vortical structures are compact, with only mild deformation. As is reduced, density contrasts sharpen and secondary vortices emerge near the outer corners of the domain. Their shape is consistent with Kelvin–Helmholtz-type growth, produced as the main vortical structures interact with outward-propagating fluctuations. The lowest-pressure case, , produces the most developed vortex system, with the largest number of secondary structures and the strongest density filamentation.
Figure 7. Density fields for the contact-driven vortex sheet test at four background pressures, . Lower pressure increases compressibility and strengthens the density filamentation associated with vortex-sheet roll-up.
The numerical solution follows this progression in all four cases. Smooth interior regions remain free of visible grid-scale noise, and the contact surfaces stay sharp throughout the evolution. Because no shocks are present, the problem gives a direct check on whether the limiting ADER-DG formulation can preserve contact layers without adding enough diffusion to erase the rolled-up vortex structure.
The same type of shear-driven roll-up appears in several astrophysical settings. Tangential velocity discontinuities at jet boundaries, and in shear-dominated regions of the interstellar medium, can undergo similar Kelvin–Helmholtz growth, producing filamentary structure and turbulent mixing.
Shock–Interface Interaction
The second problem adds the compressive forcing that is absent from the vortex-sheet test. A strong shock propagates obliquely into a stratified density field, producing a transmitted shock, a reflected wave, and a deforming material interface. In this configuration, unlike the previous one, the baroclinic term (29) is active. Across the interface, and are strongly misaligned, so the baroclinic torque deposits vorticity directly on the contact surface. This is the mechanism underlying Richtmyer–Meshkov growth [24, 25]: impulsive vorticity deposition drives the interface toward a bubble–spike morphology, which later feeds Kelvin–Helmholtz-like roll-up at the tips. Table 8 lists the initial states.
| Region | ||||
| 1 (, ) | 1.500 | 0.000 | 0.000 | 1.500 |
| 2 (, ) | 0.5323 | 1.206 | 0.000 | 0.300 |
| 3 (, ) | 0.138 | 1.206 | 1.206 | 0.029 |
| 4 (, ) | 0.5323 | 0.000 | 1.206 | 0.300 |
Table 8. Initial conditions for the shock–interface interaction. Domain , interfaces at . Non-reflecting outflow conditions on all boundaries. Resolution .
The snapshots in Figure 8 show the temporal evolution of the density field. The initial shock front moves toward the lower-left and strikes the contact interfaces obliquely. After impact, the wave separates into transmitted and reflected components, while the material interface deforms and starts to roll up. At later times the interface develops the expected Richtmyer–Meshkov morphology: heavy-fluid spikes extend into the lighter region, and lighter-fluid bubbles rise into the heavier one. Kelvin–Helmholtz-like roll-up then appears at the spike tips as secondary shear between the two fluids amplifies. As these features grow and interact, the density field becomes multiscale, combining large wave structures with smaller vortical patterns.
Figure 8. Temporal evolution of the density field in the shock–interface interaction. The snapshots show shock propagation, interface deformation, Richtmyer–Meshkov growth, and secondary Kelvin–Helmholtz roll-up.
The scheme captures this evolution without introducing spurious oscillations in the smooth regions between wave structures. The subcell limiter is triggered at shock fronts and along the deforming interface, where steep gradients form, but the high-order update is retained in smooth downstream regions. Entropy waves, visible as sharp density layers advected by the flow, remain well resolved across the domain. This behaviour reflects the intended division of labour in the a posteriori limiter: troubled zones receive a robust finite-volume update, while smooth regions continue to use the high-order DG evolution.
The same physical mechanism — baroclinic vorticity deposition at a shock-accelerated density interface — is relevant when supernova blast waves encounter dense molecular clouds in the interstellar medium. In that setting, RMI-driven mixing can help shape filamentary structures.
The one-dimensional tests of § 5 showed that the scheme resolves the three Euler wave families along a single axis. The present tests extend that assessment to multidimensional interface dynamics. The vortex-sheet problem probes contact surfaces and shear-driven roll-up without shock forcing; the shock–interface problem adds impulsive vorticity deposition, transmitted shocks, and secondary Kelvin–Helmholtz-like structure. In both cases, the ADER-DG scheme with subcell limiting and AMR handles smooth and non-smooth features within the same calculation. Related interface geometries also occur outside astrophysics, for example in atmospheric density fronts and gravity currents. The analogy is geometric and dynamical rather than one of scale, and is developed further in § 7.
Discussion: Numerical Findings and Physical Relevance
The benchmark sequence in § 5 and § 6 was designed to reduce the ambiguity of the assessment one step at a time. It tests the ADER-DG implementation in flows where smooth waves, shocks, contacts, and multidimensional interfaces appear in different combinations. The useful point is not just that each test can be run, but that the same numerical ingredients behave consistently as the character of the flow changes. For that reason, the physical interpretation is kept narrow: the results concern the inviscid, non-relativistic Euler limit, and the analogies below should be read within that limit.
Numerical Findings in Context
The five benchmark problems are intended as an escalation in physical and numerical difficulty, rather than as independent demonstrations. Sod checks the basic ordering and propagation of the Euler wave families. Shu–Osher adds a shock moving through smooth entropy structure. Woodward–Colella then asks whether the method remains robust during repeated strong-shock interactions. The vortex sheet removes shock forcing and isolates contact preservation under tangential shear. Finally, the shock–interface problem combines shocks, density interfaces, baroclinic vorticity deposition, and secondary instability growth.
Taken together, the tests support three observations. First, in one dimension the solver recovers the expected Euler wave structure: rarefactions, contacts, and shocks appear in the correct sequence, and higher polynomial order improves smooth or oscillatory regions in the way expected for high-order ADER-DG schemes [8]. Second, in strongly compressive cases, the a posteriori subcell limiter acts as a reliable fallback without noticeably washing out the smooth parts of the solution. Third, in two dimensions, the method keeps interfaces sharp enough to resolve vortical structures generated by shear and by baroclinic shock–interface interaction.
Table 9 summarises the role of each benchmark in this validation argument.
| Benchmark | Role in the validation | Physical analogue |
| Sod | Tests the ordering and propagation of rarefactions, contacts, and shocks under a strong pressure jump. | Idealised blast-wave front |
| Shu–Osher | Tests the ability of higher order to preserve post-shock oscillatory structure in an inhomogeneous flow. | Shocks in inhomogeneous gas |
| Woodward–Colella | Tests robustness under repeated strong shock interaction, reflection, and limiting. | Confined blast-wave dynamics |
| Vortex sheet | Tests contact preservation and shear-driven roll-up without shock forcing. | Jet shear layers and KHI |
| Shock–interface | Tests baroclinic vorticity deposition and shock-driven interface growth in two dimensions. | Shock–cloud interaction |
Table 9. Role of each benchmark in the validation argument.
Read together, the entries in Table 9 make the main argument of the validation. The one-dimensional tests check wave-pattern fidelity and robustness under strong compression; the two-dimensional tests show that the same ingredients remain useful when geometry, shear, and baroclinic vorticity enter. No single benchmark would be enough to support that conclusion on its own.
Astrophysical Relevance
The clearest physical connection is with idealised astrophysical hydrodynamics. During the early, non-radiative evolution of supernova remnants, the Euler equations provide the standard leading-order description of freely expanding ejecta and of the subsequent adiabatic Sedov–Taylor stage [26, 27]. The Sod and Woodward–Colella tests do not model a remnant directly, but they probe the same inviscid ingredients: strong pressure contrasts, blast-wave propagation, reflected shocks, contacts, and rarefactions. Shu–Osher adds the interaction of a shock with a pre-existing density modulation, a mechanism connected to post-shock structure generation in inhomogeneous astrophysical media and core-collapse supernova flows [28].
The shock–interface test is particularly relevant to blast waves encountering dense clouds in the interstellar medium. The present configuration is idealised: it excludes stratification, magnetic fields, radiative cooling, and realistic cloud geometry. Even so, it isolates the same leading-order inviscid process that drives Richtmyer–Meshkov growth in those systems: impulsive baroclinic vorticity deposition at a shock-accelerated density interface.
Non-relativistic astrophysical jets provide a second natural connection. A collimated jet propagating through an ambient medium contains both boundary shear layers and a shocked head region [29, 30]. The vortex-sheet benchmark isolates the first geometry, where tangential velocity jumps drive Kelvin–Helmholtz-type roll-up. The shock–interface benchmark isolates the second, where an oblique shock deforms a density interface and deposits vorticity. The results support the implementation as a starting point for idealised inviscid studies of non-relativistic jet-boundary dynamics, while stopping well short of a complete jet model.
Broader Fluid-Dynamical Analogues
These mechanisms are not uniquely astrophysical. Atmospheric density currents can be idealised locally as sharp interfaces between air masses of different density, and Kelvin–Helmholtz instability is one recognised route to clear-air turbulence in stratified shear layers [31, 32, 33]. These examples are included only to mark the broader mathematical setting of the tests. They are not additional validation cases: rotation, stratification, moisture, and external forcing are not included in the present Euler benchmarks.
Limitations
The validation remains intentionally limited. All benchmarks are one- or two-dimensional, so genuinely three-dimensional effects such as vortex stretching, turbulent cascades, and azimuthal instability modes are outside the present evidence. The governing equations are the ideal inviscid Euler equations; viscosity, magnetic fields, radiative cooling, gravity, and relativistic corrections are not included. The astrophysical and atmospheric connections above therefore apply only to idealised, inviscid, non-relativistic limits.
The assessment is also internal to this implementation. A direct comparison with established astrophysical solvers such as PLUTO [34] or FLASH [35] on the same test suite would be needed to make a competitive performance claim. In addition, this version does not report a formal multi-resolution or convergence table, nor does it quantify the fraction of cells treated by the subcell limiter. A future quantitative study should add convergence rates and troubled-cell statistics. Within these limits, the benchmark sequence supports the solver as a robust high-order Euler implementation for idealised shock- and interface-dominated flows.
Conclusions
We implemented and tested a high-order ADER-DG solver for the compressible Euler equations in ExaHyPE. The solver combines a high-order ADER-DG polynomial representation, a local space–time DG predictor, a posteriori subcell finite-volume limiting, and adaptive mesh refinement. The codes and datasets are publicly available, which should make the benchmark sequence straightforward to reproduce or extend.
The tests were chosen to build up the difficulty step by step. In one dimension they check the ordering of Euler wave families, the interaction of a shock with entropy structure, and the behaviour of the limiter under repeated strong-shock reflection. In two dimensions they add contact preservation, shear-driven roll-up, baroclinic vorticity deposition, and Richtmyer–Meshkov-type interface growth. Across this sequence, the solver behaves as intended: high-order resolution is retained in smooth parts of the flow, while the subcell limiter provides a stable fallback near shocks and steep gradients.
The most direct physical use of these tests is in idealised astrophysical hydrodynamics. Blast fronts, shock–entropy interactions, shear layers, and shock-accelerated interfaces all appear in simplified descriptions of early supernova-remnant evolution, shock–cloud interaction, and non-relativistic jet-boundary dynamics. The present results should be read in that restricted sense: they support the inviscid, non-relativistic Euler limit, not a complete model including cooling, magnetic fields, gravity, or relativistic effects.
A natural next step is a more quantitative comparison using formal multi-resolution and errors, together with statistics on troubled-cell activation. Direct comparisons with established astrophysical hydrodynamics codes on the same tests would also clarify where this implementation is most competitive. Within the scope tested here, the benchmark suite supports the solver as a robust high-order tool for idealised flows dominated by shocks, contacts, and multidimensional interfaces.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Data availability
All simulation codes and datasets produced in this study are publicly available. Repository links are provided in section A.
Code and Data Repository
The codes and datasets used for the benchmark calculations are publicly available through the resources listed in Table 10. They include analytical reference tools, predictor notebooks, and the ExaHyPE configurations used for the one- and two-dimensional Euler benchmarks discussed in § 5 and § 6.
| Resource | Language / framework | Public link |
| Analytical solution to the Sod shock tube problem | Python | https://github.com/anmsuarezma/Master_Thesis/blob/main/riemannSolution.ipynb |
| Galerkin predictor of order 3 for the linear advection equation | Python | https://github.com/anmsuarezma/Master_Thesis/blob/main/galerkin3Order.ipynb |
| Lagrange nodal points for interpolation in DG schemes | Python | https://github.com/anmsuarezma/Master_Thesis/blob/main/polynomialLagrangeNodalPoints.ipynb |
| ADER-DG Sod-type benchmark configuration using ExaHyPE | ExaHyPE | https://github.com/anmsuarezma/Master_Thesis/blob/main/sod2D%20.tar.xz |
| Numerical solution to the Shu–Osher problem | ExaHyPE | https://osf.io/5kamd |
| Woodward–Colella blast-wave simulation | ExaHyPE | https://osf.io/hbfc3 |
| Vortex-sheet dynamics with adaptive refinement | ExaHyPE | https://osf.io/kdjzb |
| Shock–interface interaction in compressible hydrodynamics | ExaHyPE | https://osf.io/2v9xd |
Table 10. Public code and data resources used in this study.
References
- [1]R. J. LeVeque. Numerical Methods for Conservation Laws. Basel: Birkhäuser, 1992. DOI: 10.1007/978-3-0348-8629-1.DOI
- [2]E. F. Toro. Riemann Solvers and Numerical Methods for Fluid Dynamics: A Practical Introduction. 3rd. Berlin: Springer, 2009. DOI: 10.1007/b79761.DOI
- [3]S. K. Godunov. “A Difference Method for Numerical Calculation of Discontinuous Solutions of the Equations of Hydrodynamics”. In: Matematicheskii Sbornik 47.89 (1959), pp. 271–306.
- [4]E. F. Toro and V. A. Titarev. “ADER: Arbitrary High Order Godunov Approach”. In: Journal of Scientific Computing 17.1–4 (2002), pp. 609–618. DOI: 10.1023/A:1015126814947.DOI
- [5]V. A. Titarev and E. F. Toro. “ADER Schemes for Three-Dimensional Non-Linear Hyperbolic Systems”. In: Journal of Computational Physics 204.2 (2005), pp. 715–736. DOI: 10.1016/j.jcp.2004.10.028.DOI
- [6]B. Cockburn and C.-W. Shu. “The Runge–Kutta Discontinuous Galerkin Method for Conservation Laws V: Multidimensional Systems”. In: Journal of Computational Physics 141.2 (1998), pp. 199–224. DOI: 10.1006/jcph.1998.5892.DOI
- [7]M. Dumbser and M. Käser. “Arbitrary High Order Non-oscillatory Finite Volume Schemes on Unstructured Meshes for Linear Hyperbolic Systems”. In: Journal of Computational Physics 221.2 (2007), pp. 693–723. DOI: 10.1016/j.jcp.2006.06.043.DOI
- [8]M. Dumbser et al. “A Unified Framework for the Construction of One-Step Finite-Volume and Discontinuous Galerkin Schemes on Unstructured Meshes”. In: Journal of Computational Physics 227.18 (2008), pp. 8209–8253. DOI: 10.1016/j.jcp.2008.05.025.DOI
- [9]M. Dumbser et al. “A Posteriori Subcell Limiting of the Discontinuous Galerkin Finite Element Method for Hyperbolic Conservation Laws”. In: Journal of Computational Physics 278 (2014), pp. 47–75. DOI: 10.1016/j.jcp.2014.08.009.DOI
- [10]O. Zanotti et al. “Space-Time Adaptive ADER Discontinuous Galerkin Finite Element Schemes with a Posteriori Sub-Cell Finite Volume Limiting”. In: Computers & Fluids 118 (2015), pp. 204–224. DOI: 10.1016/j.compfluid.2015.06.020.DOI
- [11]O. Zanotti, F. Fambri, and M. Dumbser. “Solving the Relativistic Magnetohydrodynamics Equations with ADER Discontinuous Galerkin Methods, a Posteriori Subcell Limiting and Adaptive Mesh Refinement”. In: Monthly Notices of the Royal Astronomical Society 452.3 (2015), pp. 3010–3029. DOI: 10.1093/mnras/stv1510.DOI
- [12]F. Fambri, M. Dumbser, and O. Zanotti. “Space-Time Adaptive ADER-DG Schemes for Dissipative Flows: Compressible Navier–Stokes and Resistive MHD Equations”. In: Computer Physics Communications 220 (2017), pp. 297–318. DOI: 10.1016/j.cpc.2017.08.001.DOI
- [13]I. Peshkov et al. “Continuum Mechanics and Thermodynamics in the Hamilton and the Godunov-type Formulations”. In: Continuum Mechanics and Thermodynamics 30 (2019), pp. 1343–1378. DOI: 10.1007/s00161-018-0621-2.DOI
- [14]M. Dumbser et al. “A Well-Balanced Discontinuous Galerkin Method for the First-Order Z4 Formulation of the Einstein–Euler System”. In: Journal of Computational Physics 504 (2024), p. 112875. ISSN: 0021-9991. DOI: 10.1016/j.jcp.2024.112875. URL: https://www.sciencedirect.com/science/article/pii/S0021999124001244.DOI
- [15]A. Reinarz et al. “ExaHyPE: An Engine for Parallel Dynamically Adaptive Simulations of Wave Problems”. In: Computer Physics Communications 254 (2020), p. 107251. DOI: 10.1016/j.cpc.2020.107251.DOI
- [16]K. Duru et al. “A Stable Discontinuous Galerkin Method for Linear Elastodynamics in 3D Geometrically Complex Elastic Solids Using Physics Based Numerical Fluxes”. In: Computer Methods in Applied Mechanics and Engineering 389 (Feb. 2022), p. 114386. DOI: 10.1016/j.cma.2021.114386.
- [17]G. A. Sod. “A Survey of Several Finite Difference Methods for Systems of Nonlinear Hyperbolic Conservation Laws”. In: Journal of Computational Physics 27.1 (1978), pp. 1–31. DOI: 10.1016/0021-9991(78)90023-2.DOI
- [18]C.-W. Shu and S. Osher. “Efficient Implementation of Essentially Non-Oscillatory Shock-Capturing Schemes, II”. In: Journal of Computational Physics 83.1 (1989), pp. 32–78. DOI: 10.1016/0021-9991(89)90222-2.DOI
- [19]P. Woodward and P. Colella. “The Numerical Simulation of Two-Dimensional Fluid Flow with Strong Shocks”. In: Journal of Computational Physics 54.1 (1984), pp. 115–173. DOI: 10.1016/0021-9991(84)90142-6.DOI
- [20]L. D. Landau and E. M. Lifshitz. Fluid Mechanics. 2nd. Vol. 6. Course of Theoretical Physics. Oxford: Pergamon Press, 1987.
- [21]P. D. Lax. “Hyperbolic Systems of Conservation Laws II”. In: Communications on Pure and Applied Mathematics 10.4 (1957), pp. 537–566. DOI: 10.1002/cpa.3160100406.DOI
- [22]M. Ben-Artzi and J. Falcovitz. “A Second-Order Godunov-Type Scheme for Compressible Fluid Dynamics”. In: Journal of Computational Physics 55.1 (1984), pp. 1–32. DOI: 10.1016/0021-9991(84)90013-5.
- [23]C. E. Shannon. “A Mathematical Theory of Communication”. In: Bell System Technical Journal 27.3 (1948), pp. 379–423. DOI: 10.1002/j.1538-7305.1948.tb01338.x.DOI
- [24]R. D. Richtmyer. “Taylor instability in shock acceleration of compressible fluids”. In: Communications on Pure and Applied Mathematics 13.2 (1960), pp. 297–319. DOI: 10.1002/cpa.3160130207.DOI
- [25]E. E. Meshkov. “Instability of the interface of two gases accelerated by a shock wave”. In: Fluid Dynamics 4.5 (1969), pp. 101–104. DOI: 10.1007/BF01015969.DOI
- [26]R. A. Chevalier. “Self-similar Solutions for the Interaction of Stellar Ejecta with an External Medium”. In: Astrophysical Journal 258 (1982), pp. 790–797. DOI: 10.1086/160126.DOI
- [27]J. K. Truelove and C. F. McKee. “Evolution of Nonradiative Supernova Remnants”. In: Astrophysical Journal Supplement Series 120 (1999), pp. 299–326. DOI: 10.1086/313176.DOI
- [28]D. Radice et al. “Turbulence in Core-Collapse Supernovae”. In: Journal of Physics G: Nuclear and Particle Physics 45 (Apr. 2018). DOI: 10.1088/1361-6471/aab872.DOI
- [29]P. E. Hardee. “Hydrodynamic Stability of Jets Produced by Mass Accreting Systems”. In: Astrophysical Journal 234 (1979), pp. 47–57.
- [30]M. L. Norman and K.-H. A. Winkler. “Why Ultrarelativistic Numerical Hydrodynamics Is Difficult”. In: Los Alamos Science 12 (1985), pp. 70–82.DOI
- [31]J. B. Klemp, R. Rotunno, and W. C. Skamarock. “On the Dynamics of Gravity Currents in a Channel”. In: Journal of Fluid Mechanics 269 (1994), pp. 169–198. DOI: 10.1017/S0022112094001527.DOI
- [32]J. W. Miles. “On the Stability of Heterogeneous Shear Flows”. In: Journal of Fluid Mechanics 10.4 (1961), pp. 496–508. DOI: 10.1017/S0022112061000305.DOI
- [33]D. C. Fritts and M. J. Alexander. “Gravity Wave Dynamics and Effects in the Middle Atmosphere”. In: Reviews of Geophysics 41.1 (2003), p. 1003. DOI: 10.1029/2001RG000106.DOI
- [34]A. Mignone et al. “PLUTO: A Numerical Code for Computational Astrophysics”. In: Astrophysical Journal Supplement Series 170 (2007), pp. 228–242. DOI: 10.1086/513316.DOI
- [35]B. Fryxell et al. “FLASH: An Adaptive Mesh Hydrodynamics Code for Modeling Astrophysical Thermonuclear Flashes”. In: Astrophysical Journal Supplement Series 131 (2000), pp. 273–334. DOI: 10.1086/317361.DOI