When simulating physical systems governed by hyperbolic conservation laws—such as the movement of air over an airfoil, the swirling of smoke, or the behavior of astrophysical plasma—we must project continuous partial differential equations (PDEs) onto discrete representations. Whether we slice space into a grid or break it into a collection of moving particles, this discretization introduces an unavoidable side effect: numerical dissipation.
To understand why this happens, consider a simple one-dimensional advection equation. When we approximate a spatial derivative using discrete points, we rely on a Taylor series expansion. By truncating this expansion to fit a finite grid, the leading-edge error term mathematically manifests as a higher-order derivative—specifically, a diffusion term. Consequently, the discrete algorithm ends up solving an equation that behaves as if it has an extra, artificial viscosity.
As noted in a comprehensive survey by Xiaokun Wang et al. in Physics-based fluid simulation in computer graphics: Survey, research trends, and challenges, this artificial viscosity acts as an unintended low-pass filter. It drains kinetic energy from the simulation, causing fine-scale features like small eddies, thin smoke plumes, and delicate water ripples to prematurely blur and disappear over time, as demonstrated by Hong-Xing Yu et al. in Fluid Simulation on Neural Flow Maps and by Xiaosheng Li et al. in Erratum to: Dynamic BFECC Characteristic Mapping method for fluid simulations.
The Fundamental Tradeoff: Shocks vs. Turbulence
It is tempting to think we can eliminate numerical dissipation simply by using centered, non-dissipative approximations. However, fluid simulations face a harsh mathematical reality near discontinuities like shock waves. When a fluid property jumps instantly across a single cell, high-order polynomial approximations attempt to fit this step function, resulting in severe overshoots and high-frequency oscillations known as the Gibbs phenomenon. Without some form of dissipation to dampen these spurious oscillations, the simulation quickly becomes unstable and fails.
As argued by Yue Li et al. in A low-dissipation shock-capturing framework with flexible nonlinear dissipation control, high-fidelity fluid simulation requires navigating a delicate balance: we need high numerical dissipation locally to stably capture sharp discontinuities, but low dissipation globally to resolve fine-scale turbulent fluctuations. Traditional solvers often over-damp the entire domain, smoothing out sub-grid dynamics and high-frequency physical features, a limitation highlighted by Kim Stachenfeld et al. in Learned Simulators for Turbulence.
Nonlinear Limiters and High-Order Schemes
To resolve this conflict, modern numerical methods use adaptive, nonlinear strategies that dynamically adjust dissipation based on the local behavior of the flow. Weighted Essentially Non-Oscillatory (WENO) and Targeted ENO (TENO) schemes achieve this by evaluating multiple candidate stencils for interpolation. If a stencil contains a shock, its weight is reduced to favor a more stable, dissipative stencil; if the region is smooth, the scheme combines stencils to achieve high-order accuracy with minimal dissipation.
The authors of An efficient targeted ENO scheme with local adaptive dissipation for compressible flow simulation, Yiqing Shen 0002 et al., designed a framework that dynamically modulates its dissipation parameters based on local flow smoothness, preventing the over-smearing of fine turbulent structures. In a similar vein, Arpit Babbar et al. introduced a subcell blending limiter in Multiderivative Runge-Kutta Flux Reconstruction for Hyperbolic Conservation Laws to enforce physical admissibility near shocks without degrading high-order precision in smooth zones.
Rather than merely minimizing error, schemes can also be designed to respect underlying physical invariants. Praveen Chandrashekar demonstrates in Kinetic Energy Preserving and Entropy Stable Finite Volume Schemes for Compressible Euler and Navier-Stokes Equations that isolating the mathematical mechanisms of kinetic energy preservation from entropy stability ensures that artificial dissipation only acts where thermodynamically required, preventing the unphysical draining of turbulent energy.
Bypassing the Grid with Lagrangian and Geometric Methods
An alternative approach to mitigating dissipation is to abandon fixed Eulerian grids entirely. Grid-based methods suffer from numerical diffusion primarily during the advection step, where fluid properties must be interpolated across fixed cell boundaries as the fluid moves.
As reviewed by Rio Yokota et al. in Vortex Methods for the Simulation of Turbulent Flows: Review, Lagrangian vortex methods circumvent this by tracking fluid motion using discrete particles that carry vorticity directly. Because the particles move along the flow trajectories, the method eliminates grid-bound interpolation errors and their associated numerical diffusion.
To bring these geometric advantages to hybrid particle-grid frameworks, Mohammad Sina Nabizadeh et al. developed Fluid Implicit Particles on Coadjoint Orbits. By treating fluid motion as a Hamiltonian system evolving on coadjoint orbits, their method preserves topological circulation and intricate vortical structures even at exceptionally low grid resolutions.
When dealing with boundaries, Duowen Chen et al. showed in Solid-Fluid Interaction on Particle Flow Maps that accumulating coupling forces via path integrals preserves vorticity across solid-fluid interfaces, preventing the artificial damping of wakes behind moving objects.
Data-Driven and Neural Dissipation Control
Recently, researchers have begun integrating machine learning to precisely optimize numerical schemes or replace traditional spatial discretizations entirely. Instead of relying on hand-tuned heuristics to balance the dissipation-accuracy tradeoff, neural networks can be trained on high-resolution data to predict sub-grid dynamics or correct truncation errors.
For instance, Jizhong Zhou et al. introduced a hybrid approach in Data-driven optimized high-order WENO schemes with low-dissipation and low-dispersion, using a neural network to optimize classical WENO weights. By training the network to minimize spectral errors, they achieved significantly lower dissipation and dispersion, capturing high-frequency waves that standard methods typically smear out.
For shock-dominated flows, Daniel Leibovici et al. demonstrated in FC-based shock-dynamics solver with neural-network localized artificial-viscosity assignment that a neural network can be trained to allocate artificial viscosity exclusively to the core of a shock, leaving smooth regions entirely untouched by dissipative damping.
Taking a more radical step away from classical grids, Rundi Wu proposed in Implicit Neural Spatial Representations for Time-dependent PDEs to represent the spatial domain as a continuous neural field. This approach bypasses traditional mesh discretizations entirely, offering a promising avenue for mitigating the accumulated numerical diffusion that degrades long-duration time-dependent simulations.