Active Flux: Hybrid Numerical Method

• Active Flux is a hybrid numerical scheme that uses conservative cell averages and interface point values to achieve a globally continuous reconstruction.
• The method attains third-order accuracy on Cartesian grids by leveraging Simpson quadrature and high-order finite-difference approximations for interface updates.
• AF is versatile, having been extended to compressible Euler, Navier–Stokes, and kinetic plasma systems, thereby reducing numerical dissipation in complex flows.

Active Flux (AF) is a numerical method for hyperbolic conservation laws that extends the Finite Volume method by augmenting cell averages with point values located on cell interfaces. Because these interface values are shared between adjacent cells, AF produces a globally continuous reconstruction while evolving the cell averages in a conservative manner. In its classical form, AF is third order accurate, has a compact stencil in space and time, and does not require a Riemann solver; instead, it evolves the boundary point values and uses them to compute the numerical flux by quadrature (Abgrall et al., 2023).

1. Core paradigm and historical development

AF is consistently described as a hybrid finite-volume/finite-element approach. Its defining feature is the simultaneous use of conservative cell averages and interface point values. The cell averages provide exact conservation and correct shock-capturing in the finite-volume sense, whereas the interface values supply a single-valued continuous field across cell boundaries through a continuous reconstruction (Abgrall et al., 2023).

The early AF literature emphasized a fully discrete formulation. In that setting, the point values at cell boundaries are updated by an exact or approximate evolution operator applied to the continuous reconstruction, and the cell-average update uses the evolved boundary values to construct the intercell flux by Simpson-type quadrature (Barsukow, 2020). This formulation was particularly natural for linear problems, where exact evolution operators are available and AF was shown to be structure preserving (Barsukow, 2021).

For nonlinear hyperbolic systems, exact evolution operators are rarely available. This motivated approximate evolution operators based on characteristic tracing, fixed-point iteration, and eigen-decomposition. In one dimension, third-order accurate approximate operators were developed for nonlinear scalar laws and for nonlinear systems, together with an entropy fix and a new limiting strategy (Barsukow, 2020). A parallel line of work extended AF to hyperbolic balance laws with source terms while maintaining third-order accuracy, and showed how to obtain a well-balanced or stationarity preserving method for linear acoustics with gravity (Barsukow et al., 2021).

A major later development was the semi-discrete formulation. In the semi-discrete AF method on Cartesian grids, the point values are advanced by a method-of-lines discretization of the strong form of the PDE, so nonlinear hyperbolic systems in multiple dimensions can be treated without requiring exact or approximate evolution operators. The 2023 semi-discrete formulation for the compressible Euler equations on Cartesian grids is explicitly presented as overcoming the evolution-operator bottleneck for nonlinear multidimensional problems (Abgrall et al., 2023).

2. Degrees of freedom, reconstruction, and conservative update

In one space dimension, the classical AF discretization on a cell $I_i=[x_{i-\frac12},x_{i+\frac12}]$ stores the cell average $\bar U_i$ and the two interface point values $U_{i-\frac12}$ and $U_{i+\frac12}$. These three degrees of freedom determine a unique quadratic reconstruction that is continuous across interfaces and matches the cell average (Şugar-Gabor, 2023). In two dimensions on Cartesian grids, the third-order semi-discrete method stores one cell average and eight boundary point values: four edge centers and four vertices. Inside each cell, AF reconstructs the unique biparabolic polynomial

$q_{ij}^{\mathrm{recon}}(x,y)\in Q_2\otimes Q_2$

that interpolates the eight boundary values and has cell average $\bar q_{ij}$ (Abgrall et al., 2023).

For a conservation law

$\partial_t q+\nabla\cdot f(q)=0,$

the semi-discrete cell-average update on a uniform Cartesian grid is obtained by integrating over the cell and approximating the edge integrals by 3-point Gauss-Lobatto, i.e. Simpson quadrature. In the two-dimensional Euler formulation this yields

$$\frac{d}{dt}\bar q_{ij} +\frac1{\Delta x}\sum_{K\in\{-\frac12,0,+\frac12\}}\omega_K\bigl[f^x(q_{i+\frac12,j+K})-f^x(q_{i-\frac12,j+K})\bigr] +\frac1{\Delta y}\sum_{K\in\{-\frac12,0,+\frac12\}}\omega_K\bigl[f^y(q_{i+K,j+\frac12})-f^y(q_{i+K,j-\frac12})\bigr]=0,$$

with $\omega_{-\frac12}=\omega_{+\frac12}=1/6$ and $\omega_0=2/3$ (Abgrall et al., 2023). Because Simpson’s rule is exact for parabolas and the reconstruction is exact for any biparabolic field, the method is formally third order in space (Abgrall et al., 2023).

The interface point values are advanced by discretizing the strong form

$\bar U_i$0

where $\bar U_i$1 and $\bar U_i$2. The spatial derivatives at the interfaces are replaced by one-sided high-order finite differences derived from the reconstruction, and hyperbolicity permits flux-Jacobian splitting into positive and negative parts through eigen-decomposition (Abgrall et al., 2023). This retains the AF principle that upwinding enters the point-value evolution while the cell averages remain strictly conservative.

A distinct semi-discrete AF construction was later proposed on overlapping staggered meshes. In that framework, conservative variables are evolved on the primary mesh, primitive variables are evolved on staggered meshes by a path-conservative central-upwind scheme, and the primitive solution is used to evaluate simple numerical fluxes for the conservative update. A post-processing step conservatively couples the two sets of variables and enforces nonlinear stability (Abgrall et al., 1 May 2025). This suggests that “Active Flux” denotes a broader family of schemes centered on redundant but tightly coupled solution representations.

3. Time integration, upwinding, and nonlinear stability

For the semi-discrete Cartesian AF method, the coupled ODE system for cell averages and point values is advanced by the strong-stability-preserving third-order Runge–Kutta method,

$\bar U_i$3

which matches the third-order spatial accuracy (Abgrall et al., 2023).

Nonlinear stability has been a central issue in AF research. One strategy is geometric limiting of the reconstruction. In the third-order semi-discrete Cartesian Euler method, a two-stage multidimensional limiter is applied in each cell. First, edge-wise limiting replaces a non-monotone edge parabola by a piecewise-linear “hat” function if the edge data are non-monotone or if the edge parabola would have an artificial extremum. Second, cell-wise limiting replaces the interior of an overshooting reconstruction by a constant plateau on a shrunken rectangle and smoothly blends this plateau to the edge reconstructions. The construction preserves the cell average exactly, enforces a global maximum principle, and retains global continuity because neighboring cells agree on the edge profile (Abgrall et al., 2023).

A second line of development reformulates the point-value update through flux vector splitting rather than Jacobian splitting. In one dimension, Jacobian-splitting AF methods were shown to suffer from a transonic issue for nonlinear problems due to inaccurate estimation of the upwind direction. Flux vector splitting provides a natural and uniform remedy, and in two dimensions it also addresses the mesh alignment issue that appeared in quasi-2D Riemann problems on Cartesian meshes (Duan et al., 2024). Bound-preserving AF methods were then built by rewriting the cell-average update as a convex combination of original high-order fluxes and robust low-order local Lax–Friedrichs or Rusanov fluxes, while point values are limited by a similar blending or scaling strategy. For scalar laws this enforces a maximum principle, and for the compressible Euler equations it enforces positivity of density and pressure (Duan et al., 2024).

A third stabilization mechanism is a smoothness indicator intrinsic to AF. The 2025 smoothness indicator measures the difference between two AF solution representations evolved at each time step. In smooth regions this difference is proportional to the order of the underlying AF method, whereas in “rough” parts it is $\bar U_i$4. In the reported one-dimensional Euler tests, the indicator successfully isolated discontinuities and contact surfaces while remaining vanishingly small in genuinely smooth regions (Chertock et al., 1 May 2025).

4. Accuracy, arbitrary-order constructions, and mesh extensions

The classical AF method is third order, but arbitrary-order extensions have been constructed in one dimension by three different routes. One can increase the stencil while keeping the same degrees of freedom, increase the number of point values, or include higher moments as new degrees of freedom. These routes have different properties and reflect different views of the relation of AF to the families of Finite Volume, Finite Difference, and Finite Element methods (Abgrall et al., 2022).

The first route keeps the classical cell averages and interface point values but replaces the interface update by higher-order finite-difference stencils. The second route enriches each cell with interior point values and reconstructs a higher-degree globally continuous polynomial. The third route adds higher moments and reconstructs a polynomial from moments and shared interface values (Abgrall et al., 2022). A plausible implication is that AF is best viewed not as a single fixed scheme, but as a design pattern in which continuity at interfaces is combined with conservative cell averages.

AF has also been extended beyond uniform Cartesian meshes. On triangular meshes for compressible flow problems, third- and fourth-order continuous polynomial spaces were built from vertex values, edge-midpoint values, and a cell average, with a bubble function furnishing the internal moment. The resulting method attained the expected order of precision in both linear and nonlinear cases and employed MOOD as an a posteriori fallback to first order in troubled cells (Abgrall et al., 2023). On Cartesian meshes with adaptive mesh refinement, AF was implemented in ForestClaw using ghost cells, conservative flux correction at coarse–fine interfaces, and subcycling in time; the resulting method was reported as third order accurate and conservative (Calhoun et al., 2022).

Semi-discrete AF has also moved beyond purely hyperbolic conservation laws. A fourth-order AF method for one- and two-dimensional parabolic problems rewrites the diffusion equation as a degenerate first-order system, evolves cell averages conservatively, and updates point values by fourth-order central finite differences. With SSP-RK3, the maximum CFL number for stability is $\bar U_i$5 in 1D, and positivity-preserving limitings were incorporated for the porous medium equation (Duan, 13 Oct 2025). In diffusive scaling for the hyperbolic heat equation, the Jacobian-splitting AF method was shown to be asymptotic-preserving in the sense that its limit scheme discretizes the limit heat equation (Duan et al., 7 Aug 2025).

5. Variational structure and relation to discontinuous Galerkin methods

A variational interpretation of semi-discrete AF was established by showing that the method on Cartesian meshes can be obtained from a Petrov–Galerkin formulation with a continuous trial space and a discontinuous, biorthogonal test space. The trial space is globally $\bar U_i$6 and uses shared interface degrees of freedom together with cell moments, whereas the test space is chosen so that the mass matrix is the identity. In this formulation, discontinuities in the test functions encode upwinding, and the method appears explicitly as intermediate between DG and CG (Barsukow, 20 Aug 2025).

The same work characterizes AF as CG-like because the trial space is globally continuous and uses shared nodal degrees of freedom at cell boundaries, but DG-like because the test space is discontinuous and there is no mass-matrix inversion. The moment equations recover the standard finite-volume cell-average update, while the interface equations arise from the action of the interface test functions on $\bar U_i$7 (Barsukow, 20 Aug 2025).

A stronger statement was later proved for linear problems: semi-discrete AF and DG are the same method after a linear mapping between their respective degrees of freedom. In one and several dimensions, and “in some sense” also for nonlinear problems, AF with on-average $\bar U_i$8 degrees of freedom per cell reconstructs a polynomial of degree $\bar U_i$9, whereas DG with the same number of degrees of freedom uses only polynomials of degree $U_{i-\frac12}$0. Under the AF–DG mapping, the updates of the degrees of freedom agree, and the Radau polynomials that appear in DG superconvergence become the bridge between the two methods (Barsukow et al., 18 Mar 2026). This provides a concrete explanation of why DG exhibits superconvergence at Radau points: at those points, AF “shines through” as the background high-order scheme behind DG (Barsukow et al., 18 Mar 2026).

6. Representative applications, performance, and open questions

The compressible Euler equations remain the canonical AF application. In the 2023 semi-discrete Cartesian study, a smooth Gaussian-bump initial condition showed third-order convergence in $U_{i-\frac12}$1 of the point values up to $U_{i-\frac12}$2. A radial Sod shock tube was run on $U_{i-\frac12}$3 cells; without limiting, small oscillations appeared near shocks, while with limiting they were suppressed. Four classic Lax–Liu configurations were computed on $U_{i-\frac12}$4 grids with $U_{i-\frac12}$5, and the Munz–Roller Kelvin–Helmholtz test on $U_{i-\frac12}$6 and $U_{i-\frac12}$7 grids with $U_{i-\frac12}$8 demonstrated resolution of both smooth acoustic waves and emerging vortical structures in the nearly incompressible regime without dedicated low-Mach fixes (Abgrall et al., 2023).

AF has also been developed for viscous and relaxation systems. A 2023 study on time-dependent, viscous, compressible flows presented one-dimensional evolution operators for linear and nonlinear hyperbolic conservation systems, a source-term extension, a hyperbolic formulation of the diffusion equation, and a hyperbolic reformulation of the compressible Navier–Stokes equations together with an operator-splitting approach (Şugar-Gabor, 2023). For ideal magnetohydrodynamics, a positivity-preserving AF scheme with the Godunov–Powell source term used a quadratic reconstruction, local Lax–Friedrichs flux vector splitting for the point values, and parametrized flux and scaling limiters to preserve density and pressure positivity (Duan et al., 5 Jun 2025).

In kinetic plasma simulation, AF has been combined with operator splitting. For the Vlasov–Poisson system, a split-step AF method proposed three flux-integral formulations: a one-dimensional reconstruction of second order, a third-order reconstruction based on information along each dimension, and a third-order discrepancy formulation. On 1D1V weak and strong Landau damping and two-stream instability benchmarks, the third-order flux-integral variant yielded third-order convergence and machine-exact mass conservation, and AF produced visibly sharper fine-scale structures at equal total degrees of freedom (Hensel et al., 2024). For the six-dimensional Vlasov–Maxwell system, operator splitting reduced the problem to repeated one-dimensional AF updates; the method was reported to have significantly lower dissipation and reduced anisotropy than a related semi-Lagrangian finite-volume benchmark while also offering lower computational cost (Grünwald et al., 27 Nov 2025).

Across these developments, several open questions recur. For the semi-discrete Cartesian Euler method, the limiter is explicitly described as fairly complex; a rigorous entropy-stability proof and positivity preservation of pressure and density remain subjects for future work; and extension to unstructured meshes and higher orders is possible but non-trivial (Abgrall et al., 2023). In the broader AF literature, further work is also directed toward adaptive strategies, more aggressive hybridization in flagged cells, higher-dimensional structure-preserving formulations, and efficient implementations on general meshes (Chertock et al., 1 May 2025).