---
title: Finite Volume Subcell Element Methods
url: https://www.emergentmind.com/topics/finite-volume-subcell-element-fvse
type: topic
---

# Finite Volume Subcell Element Methods

Searching arXiv for relevant papers on finite volume subcell element and closely related DG/FV subcell methods.
Finite Volume Subcell Element (FVSE) denotes a class of hybrid high-order/low-order discretization strategies in which a high-order element-local representation—most commonly a discontinuous Galerkin (DG) or DG spectral element method (DGSEM)—is coupled to a finite-volume (FV) evolution on a finer subgrid of subcells inside each element. In this framework, the element is treated as a container for a conservative subcell finite-volume dynamics, either continuously through flux blending or intermittently through troubled-cell replacement. The defining objective is to preserve the high-order accuracy and subcell resolution of DG in smooth regions while recovering the robustness of FV discretizations near shocks, positivity threats, and other nonlinear instabilities. Although the exact label “Finite Volume Subcell Element” is not used uniformly across the literature, closely aligned formulations include local subcell monolithic DG/FV schemes on unstructured grids [2407.16815], subcell limiting strategies for DGSEM on Legendre–Gauss–Lobatto (LGL) nodes [2202.00576], a posteriori MOOD-type subcell limiters for ADER-DG and \(P_NP_M\) schemes [1406.7416, 2010.04853], positivity-preserving DGSEM/FV blending [2102.06017], hybrid artificial-viscosity/FVSE shock capturing in aerospike simulations [2508.10275], and h-adaptive FV subcell shock capturing on heterogeneous curvilinear meshes [2604.27735].

## 1. Conceptual definition and scope

FVSE is best understood as an embedded finite-volume representation inside a higher-order element method. The core idea appears in several closely related forms.

In the local subcell monolithic DG/FV formulation, the DG polynomial on each cell \(\omega_c\) is reinterpreted through subcell mean values on a subdivision
\[
\omega_c = \bigcup_{m=1}^{N_s} S_m^c,
\]
with subcell means
\[
u_m^c=\frac{1}{|S_m^c|}\int_{S_m^c} u_h^c\,dV.
\]
The DG evolution is then rewritten as a conservative finite-volume-like update on this subgrid, so that the scheme is conservative at the subcell level [2407.16815]. This provides a direct FVSE-type interpretation: the “element” remains the DG cell, but the actual conservative update is expressed on internal finite-volume subcells.

In DGSEM-based work, the same structural idea is realized on LGL nodes. The nodal values \(u_{ij}\) are interpreted as subcell-centered values on a local tensor-product subcell grid, and both the high-order DGSEM operator and the low-order subcell FV operator are written in the same conservative flux-difference form,
\[
J_{ij} \dot{u}_{ij} = \frac{1}{\omega_i}\left( f_{(i-1,i)j} - f_{(i,i+1)j} \right) + \frac{1}{\omega_j}\left( f_{i(j-1,j)} - f_{i(j,j+1)} \right),
\]
thereby enabling direct convex blending [2202.00576]. This suggests that FVSE is not merely a fallback discretization, but a structural reinterpretation of the element-local DG operator itself.

A different but related interpretation arises in a posteriori subcell limiting. In those approaches, the primary solution remains a high-order DG or \(P_NP_M\) polynomial on the main mesh, but troubled elements are scattered onto a finer subgrid, recomputed there by a robust FV scheme, and then reconstructed back to the polynomial space [1406.7416, 2010.04853]. In that setting, FVSE functions as a temporary finite-volume substructure embedded inside the element.

A plausible implication is that “FVSE” is best regarded as a family of subcell finite-volume embeddings rather than a single canonical algorithm.

## 2. Algebraic structure: from DG elements to subcell finite volumes

A defining feature of FVSE-type methods is that the high-order element discretization is made compatible with a conservative subcell formulation.

For the unstructured-grid monolithic DG/FV method, the starting point is the DG polynomial
\[
u_h^c=\sum_{m=1}^{N_k} u_m^c(t)\,\sigma_m^c(\mathbf{x}), \qquad N_k=\frac{(k+1)(k+2)}{2}\ \text{in 2D},
\]
together with the weak DG equations
\[
\int_{\omega_c} \partial_t u_h^c\,\psi\,dV = \int_{\omega_c} F(u_h^c)\,\nabla\psi\,dV - \int_{\partial\omega_c} \psi\, F_n\, dS, \qquad \forall \psi\in \mathbb{P}^k(\omega_c).
\]
After quadrature, this becomes
\[
M_c U_c = \Phi_c.
\]
The crucial step is then the construction of DG reconstructed fluxes \(F_{mp}\), special subcell-face fluxes that make the finite-volume subgrid evolution algebraically equivalent to the original DG residual [2407.16815]. On faces touching neighboring cells, the reconstructed flux is simply the DG numerical flux; on interior subcell faces, it is obtained from
\[
F_c = -A_c\,L_c^{-1}\left(D_c\,P_c\,M_c^{-1}\Phi_c + B_c\right).
\]
This gives the unique set of interior subcell fluxes that reproduces the DG method exactly, independent of the particular subcell partition, as long as \(P_cP_c\) is invertible [2407.16815].

In the DGSEM setting, compatibility is achieved through a flux-differencing representation. The high-order DGSEM operator and the low-order subcell FV operator share the same stencil structure, with the high-order “fluxes” constructed recursively from volume terms, for example
\[
f_{(i,i+1)j} = - \sum_{l=0}^{i} F^{\mathrm{Vol(1)}}_{lj}.
\]
Because both operators are written on the same LGL-based subcell layout, convex mixing can be performed without interpolation to a separate mesh [2202.00576].

On curvilinear meshes, compatibility additionally requires carefully designed metric terms. The subcell metric expressions
\[
_{(i,i+1)j}  = J\vec{a}^1_{0j} + \sum_{l=0}^{i} \sum_{m=0}^{N} Q_{lm} (J\vec{a}^1)_{mj}, \qquad
_{i(j,j+1)}  = J\vec{a}^2_{i0} + \sum_{l=0}^{j} \sum_{m=0}^{N} Q_{lm} (J\vec{a}^2)_{im}
\]
satisfy the discrete subcell metric identities
\[
\frac{ _{(i-1,i)j} - _{(i,i+1)j} }{\omega_i} + \frac{ _{i(j-1,j)} - _{i(j,j+1)} }{\omega_j} = \vec{0},
\]
which guarantees local conservation and free-stream preservation [2202.00576].

This body of work establishes the central algebraic point: FVSE is not a heuristic patch applied after discretization, but a conservative subcell reformulation or compatible embedding of the element-local operator.

## 3. Flux blending, switching, and limiting strategies

The most characteristic operational mechanism in FVSE-type methods is the local combination of a high-order DG-like update and a more robust low-order FV update.

In the monolithic DG/FV scheme, each face of each subcell carries two fluxes: a high-order DG reconstructed flux and a first-order FV flux
\[
F_{mp}^{\mathrm{FV}} = F(u_m^c,u_p^v,\mathbf{n}_{mp}).
\]
The actual numerical flux is a convex blend
\[
F_{mp}=F_{mp}^{\mathrm{FV}}+\theta_{mp}\big(F_{mp}-F_{mp}^{\mathrm{FV}}\big), \qquad \theta_{mp}\in[0,1].
\]
Here \(\theta_{mp}=1\) corresponds to the pure high-order DG reconstructed flux, \(\theta_{mp}=0\) to the pure first-order FV flux, and intermediate values to a local blend [2407.16815].

In DGSEM formulations, the same idea appears at either the element scale or the subcell-interface scale. Element-wise blending uses a single coefficient,
\[
\dot{u}_{ij} = (1-\alpha)\dot{u}_{ij}^{DG} + \alpha \dot{u}_{ij}^{FV},
\]
whereas subcell-wise blending blends the interface fluxes directly,
\[
f_{(a,b)} = (1-\alpha_{(a,b)}) f^{DG}_{(a,b)} + \alpha_{(a,b)} f^{FV}_{(a,b)}.
\]
Subcell-wise blending is more localized and usually less dissipative, while element-wise blending is simpler and is the case where stronger entropy-dissipation results are currently available [2202.00576].

Troubled-cell activation is implemented in several distinct ways across the literature.

A priori activation uses indicators or sensors before the update. In DGSEM subcell limiting, one option is a modal-energy detector based on
\[
\mathbb{E} = \max\left(\frac{\hat{\epsilon}_N^2}{\sum_{j=0}^{N} \hat{\epsilon}_j^2}, \frac{\hat{\epsilon}_{N-1}^2}{\sum_{j=0}^{N-1} \hat{\epsilon}_j^2}\right),
\]
followed by a smooth mapping to \(\alpha\) [2202.00576]. In the heterogeneous-mesh DGSEM/FV scheme, switching is driven by
\[
\mathcal{I}^{(\iel)} = \max\left( \mathcal{I}_{\text{jump}(\mathbf{v}^{(\iel)})}, \ C(1-\mathcal{S}^{(\iel)}) \right),
\]
with hysteresis between lower and upper thresholds [2604.27735].

A posteriori activation computes a candidate high-order solution first and then checks admissibility. The ADER-DG MOOD limiter computes an unlimited candidate \(u_h^*(x,t^{n+1})\), applies PAD and NAD criteria, and only if the cell fails is the solution scattered to subcells and recomputed by a robust FV method [1406.7416]. The corresponding \(P_NP_M\) formulation follows the same logic with physical admissibility and a relaxed discrete maximum principle (RDMP) on subcell averages [2010.04853].

A plausible implication is that the distinction between “FVSE by blending” and “FVSE by replacement” is operational rather than conceptual: both rely on the same embedded subcell finite-volume representation.

## 4. Admissibility, positivity, and entropy control

FVSE-type methods are used because high-order DG discretizations can generate spurious oscillations, negative density or pressure, or code crashes in under-resolved or shock-dominated regimes. The subcell finite-volume component provides a robust admissible state toward which the solution can be pulled.

In the positivity-preserving DGSEM limiter for the Euler equations, the blended semi-discrete operator is
\[
\dot{\mathbf{u}} = (1-\alpha)\,\dot{\mathbf{u}}^{DG} + \alpha\,\dot{\mathbf{u}}^{FV}, \qquad \alpha\in[0,1].
\]
The limiter enforces
\[
\rho \ge \beta\,\rho_{\mathrm{safe}}, \qquad p \ge \beta\,p_{\mathrm{safe}}, \qquad \beta\in(0,1],
\]
relative to a safe FV stage state obtained with \(\alpha=1\) [2102.06017]. The paper uses \(\beta=0.1\) in the numerical tests, and the correction is applied at the end of each RK stage. Density correction is algebraic, whereas pressure correction requires solving
\[
g(\alpha_{\mathrm{new}}) = p(\mathbf{u}_{\mathrm{new}}(\alpha_{\mathrm{new}})) - \beta p_{\mathrm{safe}} =0
\]
with Newton’s method [2102.06017]. This is a strictly local, conservative fallback toward a positivity-safe subcell FV update.

The MOOD-style ADER-DG limiter also combines physical admissibility detection with a relaxed discrete maximum principle. For Euler, PAD requires
\[
\rho > 0, \qquad p > 0,
\]
while NAD is imposed through
\[
\min_{y\in \mathcal{V}_i} u_h(y,t^n) - \boldsymbol{\delta} \le u_h^*(x,t^{n+1}) \le \max_{y\in \mathcal{V}_i} u_h(y,t^n) + \boldsymbol{\delta},
\]
with
\[
\boldsymbol{\delta} = \epsilon\left( \max_{y\in \mathcal{V}_i} u_h(y,t^n) -\min_{y\in \mathcal{V}_i} u_h(y,t^n) \right), \qquad \epsilon=10^{-3}
\]
[1406.7416]. The \(P_NP_M\) extension uses RDMP on subcell averages with
\[
\delta=\max\!\left(\delta_0,\, \epsilon\left[ \max_{m\in\mathcal V(i)}\max_{\beta}\mathbf v_{m,\beta}^n - \min_{m\in\mathcal V(i)}\min_{\beta}\mathbf v_{m,\beta}^n \right]\right),
\]
where \(\delta_0=10^{-5}\) and \(\epsilon=10^{-4}\) [2010.04853].

Entropy control has been analyzed most explicitly in the monolithic subcell DG/FV framework. Three levels are distinguished [2407.16815].

1. Discrete subcell entropy inequality for any entropy gives the strongest guarantee, but forces the method toward first order.

2. Semi-discrete subcell entropy stability for one entropy improves accuracy to about second order, but does not preserve arbitrary high order.

3. Semi-discrete cell entropy stability for one entropy enforces entropy inequality only on the whole cell \(\omega_c\), leading to a continuous knapsack problem
   \[
   C_c\Theta_c\le D_c,
   \]
   solved greedily, and preserving high-order accuracy in smooth regions because
   \[
   C_c\mathbf{1}-D_c = |\omega_c|\,O(h_c^{k+1}),
   \]
   so the blended scheme reduces to pure DG up to
   \[
   F_{mp}=F_{mp}^{\mathrm{DG}}+O(h_c^{k+2}).
   \]
   The paper also notes that this high-order-preserving entropy option requires modified FV fluxes evaluated on entropy-variable states and may destroy other properties such as positivity [2407.16815].

The literature therefore treats positivity, invariant-domain preservation, maximum principles, and entropy stability as related but nonidentical admissibility targets. One common misconception is that “entropy stable” automatically means positivity preserving; the cited work explicitly separates these properties [2102.06017, 2407.16815].

## 5. Subcell resolution, time stepping, and reconstruction

A recurrent design problem in FVSE-type schemes is how many subcells to place inside each element. The dominant choice in the a posteriori subcell-limiting literature is \(2N+1\) subcells per spatial dimension.

In the ADER-DG subcell limiter, troubled cells are scattered onto \((N_s)^d\) Cartesian subcells, with
\[
N_s=2N+1.
\]
This is described as optimal because it matches the maximum admissible time step of the subcell FV scheme to the maximum admissible time step of the DG scheme:
\[
\Delta t \le \frac{1}{d}\frac{1}{(2N+1)}\frac{h}{|\lambda_{\max}|}
\]
for DG and
\[
\Delta t \le \frac{1}{d}\frac{1}{N_s}\frac{h}{|\lambda_{\max}|}
\]
for the subcell FV method [1406.7416]. The same rationale appears in the \(P_NP_M\) limiter, where the subgrid contains
\[
N_\omega=(2N+1)^d
\]
subcells, chosen precisely so that the subcell FV scheme can run with the same timestep already admissible for the main discretization [2010.04853].

The heterogeneous-mesh DGSEM/FV scheme allows an arbitrary subcell resolution satisfying
\[
(\ppn+1)^d \le N_k \le (2\ppn+1)^d,
\]
but adopts
\[
N_k = (2\ppn+1)^d
\]
in the reported work [2604.27735]. The paper states that at least \((\ppn+1)^d\) subcells are needed for robust and accurate shock capturing, and the upper bound is used to avoid a stronger CFL restriction than DG.

Projection and reconstruction between the high-order element representation and the subcell FV state are equally central. In ADER-DG, scattering is performed by exact subcell averaging,
\[
v_{i,j}^n = \frac{1}{|S_{i,j}|}\int_{S_{i,j}} u_h(x,t^n)\,dx,
\]
defining the projection operator \(\mathcal{P}\), and reconstruction is defined so that the reconstructed polynomial matches the subcell averages and preserves the whole-cell integral. When \(N_s\ge N+1\), the reconstruction is well posed and
\[
\mathcal{R}\circ\mathcal{P}=\mathcal{I}
\]
[1406.7416].

In the \(P_NP_M\) setting, projection and reconstruction satisfy
\[
\mathcal P\circ\mathcal R=\mathcal I,
\]
and troubled cells may be recomputed with either second-order MUSCL-Hancock TVD FV or ADER-WENO FV on the subgrid [2010.04853]. In the heterogeneous-mesh method, DG-to-FV and FV-to-DG transfer operators are written explicitly with conservation constraints, using least squares for reconstruction [2604.27735].

This suggests that subcell resolution is not only a geometric refinement device; it is also an algebraic mechanism for retaining the information content of the element-local polynomial while enabling robust low-order evolution.

## 6. Variants, applications, and practical regimes

The FVSE concept spans several numerical roles, from shock capturing to invariant-domain enforcement and hybrid stabilization.

### Representative variants

| Variant | Operational mode | Representative source |
|---|---|---|
| Monolithic local DG/FV on subcells | Continuous convex blending of DG reconstructed fluxes and FV fluxes on each subcell face | [2407.16815] |
| DGSEM subcell limiting on LGL nodes | Element-wise or subcell-wise convex blending with compatible FV subcell operators | [2202.00576] |
| A posteriori MOOD-type subcell limiter | Troubled-cell recomputation on subcells, then reconstruction back to DG/\(P_NP_M\) | [1406.7416], [2010.04853] |
| Positivity-preserving DGSEM/FV blend | Element-local a posteriori blending toward a positivity-safe FV state | [2102.06017] |
| Hybrid AV+FVSE shock capturing | AV as primary stabilizer, FVSE as localized fallback “parachute” | [2508.10275] |
| h-adaptive mixed-element FV subcells | DGSEM/FV switching on hexahedra, prisms, tetrahedra, and pyramids | [2604.27735] |

In practical shock capturing, the aerospike-nozzle solver uses a hybrid artificial-viscosity/FVSE strategy. There, artificial viscosity is the primary stabilizer and FVSE is invoked only where AV cannot sufficiently smooth the solution gradients. The FVSE residual is blended element-wise,
\[
\mathcal{R}_{ijk} = \alpha \mathcal{R}^{FVSE}_{ijk} + (1-\alpha)\mathcal{R}^{DGSEM}_{ijk},
\]
with a low-order finite-volume residual based on Lax–Friedrichs fluxes between subcells associated with DG nodal points [2508.10275]. The paper explicitly describes FVSE as a “parachute” in the context of numerical stability. It also emphasizes that broad use of FVSE is highly dissipative: a constant \(\alpha=0.1\) everywhere suppresses major wake dynamics, whereas the preferred “CM5” configuration \((C_\mu, C_\kappa, C_m)=(0.5,0.5,0.5)\) uses FVSE only in extreme-gradient regions and captures the unsteady wake while preserving stability [2508.10275].

In the positivity-preserving DGSEM limiter, the FV subcell method is not primarily a shock-capturing device, but a positivity-safe correction mechanism for density and pressure in under-resolved vortex-dominated and shock-dominated Euler flows [2102.06017].

In the heterogeneous-mesh DGSEM/FV method, the important extension is topological generality. The mesh may contain hexahedra, prisms, tetrahedra, and pyramids, with non-hexahedral elements handled via collapsed coordinate transformations. The FV operator is a second-order TVD finite-volume scheme with generalized minmod or Barth–Jesperson limiting, and the paper verifies conservation, spatial convergence, and shock-capturing capability on curved mixed meshes, including a NACA 0012 airfoil case using a mesh with 1168 prisms and 12032 hexahedra [2604.27735].

Across these applications, the same practical pattern recurs: DG or DGSEM provides the high-fidelity baseline in smooth flow, while FVSE is activated locally to stabilize shocks, strong gradients, positivity threats, or difficult startup transients.

## 7. Relation to neighboring methods and major interpretive issues

FVSE belongs to the broader family of DG/FV hybridizations, but it is distinct from several nearby ideas.

It is not simply classical artificial viscosity. Artificial viscosity adds local diffusion inside the element and can degrade sub-element resolution; the cited mixed-mesh DGSEM/FV work contrasts this explicitly with subcell limiting, which preserves sub-element resolution better by redistributing information onto a fine internal FV grid [2604.27735]. The aerospike solver likewise treats AV and FVSE as complementary rather than interchangeable: AV is computationally efficient and less intrusive in smooth regions, whereas FVSE is a more dissipative fallback [2508.10275].

It is not standard finite volume. In all cited formulations, the method is not globally first-order FV. Rather, it retains a high-order DG or DGSEM representation almost everywhere and uses the subcell FV structure locally, either continuously or intermittently [2202.00576, 2102.06017].

It is not identical to classical finite volume element methods in the older FEM sense. The literature summarized here uses “finite volume subcell element” to denote a high-order element method with an internal finite-volume substructure, not a separate finite-volume discretization on arbitrary control volumes. This distinction is especially clear in the DGSEM-based work, where the FV scheme is explicitly co-located with the DG nodal grid [2102.06017, 2202.00576].

Several interpretive issues recur in the literature.

One is the tradeoff between localization and theory. Subcell-wise blending is more localized and less dissipative than element-wise blending, but a full entropy-stability proof for subcell-wise blending is still missing in the DGSEM framework [2202.00576]. Another is the tradeoff between entropy control and high-order accuracy. The monolithic unstructured-grid analysis shows that the strongest entropy condition drives the method essentially to first order, whereas cell-level semi-discrete entropy stability for one entropy can preserve high order but at increased complexity and possible loss of positivity [2407.16815].

A further issue concerns whether FVSE should be viewed primarily as a limiter or as a reformulation. The unstructured-grid monolithic scheme and DGSEM flux-differencing formulations support the latter view, because the high-order operator itself is recast as a subcell finite-volume update [2407.16815, 2202.00576]. The MOOD-type and \(P_NP_M\) approaches emphasize the former, because subcells are activated only in troubled regions [1406.7416, 2010.04853]. A plausible synthesis is that FVSE is both: an algebraic reinterpretation of element-local high-order methods and a practical limiting architecture built on that reinterpretation.

In current usage, FVSE therefore names a robust and technically flexible paradigm for embedding finite-volume subcell dynamics inside high-order element methods. Its enduring significance lies in precisely this dual capacity: preserving DG-level accuracy and subcell resolution where the solution is smooth, while restoring finite-volume robustness where nonlinear dynamics would otherwise make the high-order scheme fail.

Source: https://www.emergentmind.com/topics/finite-volume-subcell-element-fvse