Localized Shock Capturing

• Localized Shock Capturing is a numerical strategy that confines artificial dissipation, ensuring stability near shocks while maintaining high-order accuracy in smooth regions.
• It employs precise sensors such as modal decay indicators and neural networks to detect discontinuities and trigger localized limiting or viscosity.
• Applications span high-order DG, FV, and FE discretizations in compressible dynamics, effectively resolving shocks, turbulence, and detonation fronts.

A localized shock capturing method is a class of numerical strategies for hyperbolic conservation laws that suppress nonphysical oscillations and ensure stability in the presence of discontinuities, while strictly confining artificial dissipation to a minimal region near the shock. The concept of "localization" refers both to the spatial support of the dissipative mechanism (cell-local, element-local, or band-limited), and to property detection through neighboring or elementwise information, so that high-order accuracy is retained in smooth parts of the solution. Such methods are central to high-order discontinuous Galerkin (DG), finite volume (FV), and finite element (FE) discretizations for compressible gas dynamics, MHD, multiphase flows, and general nonlinear systems. Localized shock capturing is essential for effective simulation of strong shocks, turbulence-shock interactions, detonation fronts, and sharp interfaces on unstructured or adaptive meshes.

1. Localized Shock Capturing: Principles and Definitions

Localized shock capturing comprises two core components: (i) a sensitive detector that identifies troubled elements or regions containing discontinuities, and (ii) a dissipation or limiting mechanism—artificial viscosity, flux limiting, modal damping, subcell blending, or similar—that is applied only where needed. Unlike global or purely monotonicity-based methods, localized approaches rely on carefully chosen sensors and localization operators constructed to act only on a small set of elements surrounding the discontinuity, thereby preserving formal high-order convergence in smooth zones (Moe et al., 2015, Wong et al., 2017, Jain et al., 2023, Klöckner et al., 2011). Localization strategies vary in spatial granularity—from strictly element-neighbor stencils to vertex-star averaging, subcell FV partitions, or neural-network-based per-element shock localization.

The imposition of locality is realized through:

• Shock/regularity sensors (modal decay, TVD/WENO smoothness indicators, troubled-cell flags, entropy residuals, neural networks)
• Dissipative operators (artificial viscosity, convex blends of low/high-order schemes, rescaling of high-order corrections, nonlinear limiting, positive-definite averaging)
• Data locality (nearest-neighbor, vertex-star, or in some cases, communication-free cell processing)

A formal localized shock-capturing scheme is defined as one in which the support of added dissipation is strictly restricted by the troubled-cell detection, and outside of which the underlying high-order scheme is unmodified.

2. DG Finite Element Localized Limiters

A prototypical fully localized limiter for DG methods was introduced by Moe, Rossmanith, and Seal (Moe et al., 2015). The algorithm is elementwise and enforces a local maximum principle by comparing the high-order polynomial expansion in each element with sampled extrema in immediate neighbors. For a scalar conservation law,

• The DG field in each element is expanded as $u^h(x) = \bar u + \delta u(x)$, separating cell average and high-order modes.
• A set of sample points $\chi_i$ in the element and its neighbors is used to find local min/max in each element and among neighbors.
• Upper and lower admissible bounds $U_i^{\min}, U_i^{\max}$ are computed using a small tolerance $\alpha(h)$; high-order correction $\delta u(x)$ is rescaled by a coefficient $\theta_i \in [0,1]$ so that the field remains within these bounds.
• The limiter is mass conservative and element-local. In smooth regions, $\theta_i \to 1$, so the method is high-order-accurate.
• The method works for systems by constructing the limiter on primitive variables but applying the rescaling to the conserved basis coefficients to preserve conservation.
• The limiting depends only on neighboring cells and is parameterized by a small, scale-adaptive tolerance $\alpha(h)=C h^{r}$, recommended as $C=\mathcal O(10^2\!-\!10^3)$, $r\approx 1.5$.

Test cases in 1D and 2D (shock-entropy, double Mach, forward-facing step) validate that the approach suppresses Gibbs oscillations and achieves sharp shock profiles and high-order convergence where smooth (Moe et al., 2015).

3. Localized Artificial Dissipation and Modal Filtering Approaches

Many modern shock-capturing frameworks superimpose artificial viscosity or modal spectral viscosity exclusively in those regions marked as non-smooth. For example:

• Legendre Spectral Viscosity (LSV) (Sousa et al., 2021) applies a modal dissipation operator in polynomial spectral space, with viscosity coefficients computed per degree based on localized high-mode energy. The viscous operator is only activated in cells where a local indicator $I = \sum_{l=N-M}^N |\hat u_{l}|^2$ exceeds a threshold.
• The Quasi-Spectral Viscosity (QSV) method (Sousa et al., 2021) employs residuals of high-order filters to estimate subgrid energy in physical space and applies dissipation only where energy is detected near the spectral cut-off, confining viscosity to shock-containing regions.
• Physics-based approaches (Fernandez et al., 2018) introduce shock/thermal/shear sensors per cell to activate only the necessary artificial transport coefficients to control the local Péclet number, with strong localization via mesh- and orientation-sensitive metrics.

Localized modal-based approaches guarantee that smooth regions are unaffected and the nonphysical artifacts of excessive global smoothing are avoided.

4. Hybrid Localized Schemes: Blending and Subcell Adaptivity

Advanced localized shock-capturing approaches deploy blending of approximation spaces or subcell adaptation within elements:

• Sub-element blending (Markert et al., 2020) interprets high-order polynomial data on a cell as both a fine-grid FV subcell field and as high-order DG. Troubled sub-elements (as determined by L1-norms of reconstructed slopes) are locally updated with FV or lower-order accurate DG, while the rest of the cell employs the high-order polynomial. A single conservative common surface flux ensures conservation across heterogeneous approximation orders.
• In (Vedral et al., 23 Jan 2026), Hermite WENO information is aggregated via cell-vertex stencils, allowing the reconstruction to account for both cell and vertex-neighbor behavior, leading to robust detection and localized application of artificial viscosity. Furthermore, the artificial viscosity is non-uniformly distributed within a cell via a WENO-weighted quadrature, concentrating dissipation strictly near subcell discontinuity regions.

These blended methods can recover high-order rates globally while retaining shock resolution at the sub-element level.

5. Localized Limiting with Neural Network and Data-Driven Sensors

Recently, machine-learning-based approaches have been introduced for high-resolution localization:

• A convolutional neural network (Holistically-Nested Edge Detection type) is trained to output element-local binary maps of shock location from DG solution data (Beck et al., 2020). The detected edge map is then used to flag troubled cells and to restrict the application of subcell TVD/FV updates, slope limiting, or artificial viscosity exclusively to flagged subcells.
• In a spectral context, a neural network trained on high-order stencils classifies the local regularity class (discontinuous, $C^0$, $C^1$, $C^2$) and assigns localized viscosity with smooth windowing only over those regions where discontinuity is inferred (Bruno et al., 2021).

Data-driven sensors enable highly refined, geometry-independent localization, maintaining robustness across mesh levels and problem setups.

6. Algorithmic Workflows and Practical Implementation

Localized shock capturing is realized via procedures with these common elements:

1. For each element or node, compute a localized smoothness or discontinuity indicator (modal decay, WENO smoothness, fault detection, trained neural net, energy residual, etc.).
2. If the indicator exceeds the threshold:
   • Apply a limiter, artificial viscosity, or blending with a lower-order method only in these flagged cells or within local subcell regions.
   • Else, leave the high-order scheme unmodified.
3. For artificial viscosity, ensure it is smoothly distributed to avoid nonphysical jumps (vertex-max smoothing, P1 interpolation, or subcell windowing as needed) (Klöckner et al., 2011, Moe et al., 2015, Vedral et al., 23 Jan 2026).
4. Advance the solution using a strong stability preserving high-order time integration (often a variant of Runge–Kutta).
5. For systems (e.g., Euler), apply the shock sensor to primitive variables, but rescale or limit the conserved variables to maintain conservation (Moe et al., 2015).

A selection of localized shock-capturing workflows is shown in the table below.

Method/Class	Locality	Dissipation Type	Sensor/Trigger
(Moe et al., 2015)	1-ring cells	High-mode rescaling	Neighbor max/min
(Sousa et al., 2021)	Element	Spectral viscosity	Modal high-energy
(Markert et al., 2020)	Sub-element	FV/DG blending	L1-norm of slope diff
(Beck et al., 2020)/(Bruno et al., 2021)	Element/subcell	Local AV, FV	Data-driven (NN/CNN)
(Jain et al., 2023)	Cell/face	Local AV/LAD	Divergence/vorticity sensors
(Wong et al., 2017, Vedral et al., 23 Jan 2026)	Node/vertex	Hybrid WENO AV	WENO/CWENO/HWENO

7. Applications and Numerical Performance

Localized shock capturing has been validated on a suite of canonical and challenging test cases:

• 1D and 2D Riemann problems (e.g., shock tubes, RP1/RP3) (Moe et al., 2015, Vedral et al., 23 Jan 2026)
• Shock-entropy wave, double Mach reflection, forward-facing step (Moe et al., 2015, Sousa et al., 2021, Markert et al., 2020)
• Curved material interfaces, blast waves, multiphase flow (Wong et al., 2017, Vedral et al., 23 Jan 2026)
• Meshless irregular grids with fault-localized viscosity (Bracco et al., 2024)
• Strong detonation and stiff reaction fronts (Deng et al., 2017)
• Realistic turbulent/LES scenarios with shock-turbulence interaction (Fernandez et al., 2018, Sousa et al., 2021)

Key performance metrics include:

• Shock thickness confined to 1-3 elements/cells in high-order methods provided $p\gtrsim 4$ (Sousa et al., 2021, Vedral et al., 23 Jan 2026).
• High-order convergence in smooth regions is preserved provided the localization threshold does not trigger false positives (Moe et al., 2015, Klöckner et al., 2011).
• Localized blending and subcell adaptation ensure robust shock capturing in extreme regimes (e.g., Sedov blast, Mach 3–10) (Markert et al., 2020).

A notable conclusion is that localization of shock-capturing dissipation is essential for both stability and accuracy in modern high-order discretizations, irrespective of the choice of underlying spatial or temporal method. Representative methods achieve sharp, oscillation-free shock resolution, maintain high-order accuracy, and have minimal or zero dissipation outside discontinuities.

---

References:

(Moe et al., 2015, Klöckner et al., 2011, Markert et al., 2020, Vedral et al., 23 Jan 2026, Beck et al., 2020, Sousa et al., 2021, Sousa et al., 2021, Jain et al., 2023, Fernandez et al., 2018, Wong et al., 2017, Bracco et al., 2024, Deng et al., 2017)