When simulating continuous physical systems—such as turbulent fluids, propagating waves, or stochastic conservation laws—high-order numerical methods offer exceptional accuracy. However, they suffer from a shared vulnerability: the unphysical accumulation of energy at the smallest resolved scales.
In fluid dynamics, this manifests as spectral blocking, where energy cascades down to the grid scale but has no way to dissipate, leading to catastrophic numerical blow-ups. In spline-based methods like Isogeometric Analysis (IGA), it appears as spectral outliers—spurious high-frequency eigenvalues that drastically restrict the stable time step size. Traditional fixes, like adding uniform artificial viscosity, solve the stability problem but ruin the accuracy of the simulation by over-dampening the physically vital low-frequency modes.
High-order spectral filtering provides a surgical alternative: it selectively dampens high-frequency noise and outliers while preserving low-frequency energy and variance.
Selective Damping via Spectral Viscosity
How do we stabilize a simulation without destroying its high-order accuracy? The key is to apply dissipation exclusively where the numerical noise resides.
Lanthaler et al. demonstrated this rigorously for the 2D incompressible Euler equations in On the Convergence of the Spectral Viscosity Method for the Two-Dimensional Incompressible Euler Equations with Rough Initial Data. Instead of applying diffusion uniformly across all scales, the Spectral Viscosity (SV) method activates artificial viscosity only for high-frequency Fourier modes. This targeted diffusion maintains computational stability against non-physical oscillations while leaving smooth flow regions completely unaffected.
To push this further in extreme scenarios, such as the vanishing viscosity limit of turbulent flows, Mishra et al. leveraged a spectral hyper-viscosity technique in Efficient Computation of Large-Scale Statistical Solutions to Incompressible Fluid Flows. By using higher-order differential operators for the dissipative term, the damping effect becomes even more localized to the absolute tail of the spectrum, ensuring that the large-scale statistical solutions remain pristine and physically accurate.
Similarly, Tritschler et al. showed in On the Richtmyer–Meshkov instability evolving from a deterministic multimode planar interface that high-order compact difference schemes can be regularized using artificial fluid properties that act primarily at these microscopic scales, smoothing out gradient quantities like enstrophy without degrading macroscopic structures.
Eliminating Boundary-Driven Outliers
Often, the high-frequency artifacts that destabilize a simulation do not originate in the interior of the domain, but rather at its boundaries. Numerical models frequently fail to strictly enforce higher-order consistency conditions at the domain edges. This mismatch creates spurious high-frequency eigenvalues—spectral outliers—that artificially increase the system's numerical stiffness.
To combat this, Deng et al. introduced boundary penalty terms in Outlier removal for isogeometric spectral approximation with the optimally-blended quadratures. By weakly enforcing higher-order consistency conditions at the domain edges, this mechanism forces the model to respect the underlying physics at the boundaries, effectively suppressing non-physical outliers.
In a follow-up work on wave propagation, A boundary-penalized isogeometric analysis for second-order hyperbolic equations, the authors demonstrated that this boundary-targeted derivative penalization decouples numerical stability from physical stiffness. Because the unphysical high-frequency modes are eliminated at the source, researchers can double the stable explicit time step size without introducing excessive numerical dissipation into the primary physical waves.
Inviscid Regularization and Scale-Aware Filtering
Can we suppress high-frequency noise without adding any artificial dissipation at all? In some frameworks, the answer is yes, achieved by structurally modifying the governing equations to alter how scales interact.
Nadiga explored this in Scaling properties of an inviscid mean-motion fluid model using the Euler-$\alpha$ model. Here, a Helmholtz operator is used to relate the advection velocity to the advected field, effectively filtering the velocity field at a specific length scale. This mechanism ensures that for scales smaller than the filter length, the interactions between eddies are de-tuned. The result is a significantly steeper energy spectrum at small scales that prevents spectral blocking entirely, all while remaining within a conservative, purely inviscid Hamiltonian framework.
This concept of scale-aware filtering is also vital when integrating data-driven components into physical models. For instance, Perezhogin et al. stabilized a machine-learned ocean eddy parameterization in A stable implementation of a data-driven scale-aware mesoscale parameterization by applying a low-pass filter to the predicted stress tensor. This removes high-frequency grid-scale noise that would otherwise trigger simulation crashes near complex topography.
In computer animation, You et al. designed an adaptive time integration scheme in Energy-Controllable Time Integration for Elastodynamic Contact that naturally prioritizes low-frequency energy preservation while letting high-frequency numerical noise dissipate, keeping large-scale motions lively and bouncy while suppressing microscopic instabilities.
Preserving Realizability near Discontinuities
When simulations encounter sharp discontinuities or shocks, high-frequency oscillations manifest as "wiggles" or "ringing" instabilities. In uncertainty quantification and stochastic expansions, these oscillations can cause the solution to lose realizability—meaning the model predicts non-physical values like negative density or pressure.
Alldredge et al. addressed this in A Realizable Filtered Intrusive Polynomial Moment Method by employing high-order spectral filters combined with regularized optimization. By selectively dampening high-frequency spurious oscillations near the shock while preserving variance in smooth regions, they achieved sharp shock representation without over-dampening the statistical integrity of the simulation.
Similarly, Kusch et al. applied filtering in Dynamical low-rank approximation for Burgers' equation with uncertainty to dampen high-order moments in a stochastic expansion, mitigating oscillations near shocks and ensuring the numerical solution remains physically consistent and stable.
Key Foundational Work
Scaling properties of an inviscid mean-motion fluid model (2000) — contributed 5 concept elements to this direction.
A Subgrid-Scale Deconvolution Approach for Shock Capturing (2002) — contributed 2 concept elements to this direction.