Velocity Parameterization

• Velocity parameterization is the process of representing velocity fields using structured forms such as basis expansions, neural networks, and adaptive discretizations.
• It boosts modeling fidelity and efficiency by integrating physical constraints and reducing computational complexity through tailored analytic or data-driven approaches.
• Its applications span geophysical fluid dynamics, turbulence modeling, and robotics, with empirical validations showing reduced errors and improved conservation properties.

Velocity parameterization refers to the representation, modeling, and encoding of velocity fields, profiles, or effects through structured forms, functional bases, neural networks, or physics-informed constraints, with the aim of enhancing modeling fidelity, computational efficiency, or interpretability in complex dynamical systems. This concept underpins a broad range of domains, from geophysical fluid dynamics to rarefied gas simulations, turbulence modeling, PDE-based registration, and beyond. Approaches to velocity parameterization span traditional analytic forms, machine learning surrogates, reduced-order modeling, and discretization frameworks, each providing tailored solutions to the specific demands of high-dimensional, multiscale, or stochastic velocity phenomena.

1. Mathematical Foundations and Formal Definitions

Velocity parameterization manifests as the specification of velocity (or related fields such as subgrid momentum or distribution functions) in a reduced or constrained form suitable for estimation, closure, or simulation. Formally, this may involve:

• Expansion in a finite-dimensional basis:

$$v(x; \eta) = c_0(x) + \sum_{\ell=1}^N \eta_\ell \phi_\ell(x)$$

where $\{\phi_\ell(x)\}$ are basis functions (e.g., Gaussians, Fourier modes, Legendre polynomials), and the parameter vector $\eta$ controls the velocity variation (Mamonov et al., 2022, Hernandez, 2018).

• Neural-network closures: Convolutional neural networks (CNNs) or multi-layer perceptrons (MLPs) parameterize $\tau_i = S_i(\bar{u})$ or air-entrainment $Va$, where the mapping from resolved variables to subgrid velocity is learned from data under explicit physical and conservation constraints (Wang et al., 2024, Zhou et al., 3 Feb 2026).
• Functional splitting or adaptivity: Velocity parameterizations are sometimes embedded in frameworks that separate resolved and subgrid contributions (e.g., GEOMETRIC/GM splitting (Mak et al., 2023)), or adapt the velocity discretization to local non-equilibrium (e.g., adaptive velocity space in DVM/UGKS (Chen et al., 2022)).
• Geometric or tensor structures: The velocity direction and magnitude may be parameterized separately (e.g., speed-direction in turbulence (Olshanskii, 2020), or quaternion-based translational parameterization (Miller et al., 2020)), or transformed by auxiliary fields encoding physical constraints (e.g., Jacobian and curl for diffeomorphic image registration (Sheikhjafari et al., 2022)).

2. Velocity Parameterization in Geophysical Fluid Dynamics

Parameterization of velocity is critical in climate and ocean models due to the ubiquity of unresolved subgrid processes:

• CNN-based ocean momentum closure:

Subgrid eddy momentum fluxes $\tau_i$ are modeled as neural network surrogates $S_i(\bar{u})$, trained on high-resolution reanalysis and integrated with explicit momentum conservation constraints. The CNN is architected as a five-layer convolutional stack, accepting input velocity fields and outputting per-location momentum fixes, with both “soft” (penalty) and “hard” (mean-subtraction) global momentum conservation (Wang et al., 2024).

• Eddy-induced velocity in mesoscale parameterization:

In the GEOMETRIC GM variant, the bolus velocity $$\mathbf{u}^* = \nabla \times (\mathbf{e}_z \times \kappa \mathbf{s})$$ captures the net effect of baroclinic eddies on large-scale tracer transport. The parameterized diffusivity $\kappa$ is dynamically linked to a prognostic eddy energy budget, and a scale-separation ("splitting") approach ensures energetic consistency and scale-awareness as model resolution is varied (Mak et al., 2023).

• Wave breaking air-entrainment:

Global machine learning parameterizations (e.g., $Va^{ML}$) surrogate expensive spectral formulations and enable accurate, regime-dependent closure for air-sea fluxes. ML models are trained to map seven physical predictors (wind, wave height, age, steepness, direction, depth) to empirical $Va$, outperforming traditional bulk formulas in skill and reducing systematic biases in flux estimation (Zhou et al., 3 Feb 2026).

3. Reduced-Order, Adaptive, and Statistical Parametrizations

• Wave equation and full waveform inversion:

The velocity field $v(x)$ is expanded in a low-dimensional basis, and the forward map—from $v(x; \eta)$ to data—is replaced by projection onto a reduced-order model (ROM), with the misfit between data and model ROM being minimized. This formulation yields convexity in the inversion objective and robustness to local minima ("cycle skipping"), outperforming conventional nonlinear least squares (Mamonov et al., 2022).

• Adaptive velocity space for rarefied flows:

The discrete velocity space (DVS) in Boltzmann solvers is parameterized adaptively via tree-based refinement/coarsening criteria based on local mass and energy fractions, obviating the need for fixed grids and reducing computational complexity by an order of magnitude while ensuring conservation and accuracy (Chen et al., 2022).

• Transport-velocity formulation in SPH:

A velocity parameterization is implemented as an $O(h^2)$ displacement correction proportional to local smoothing length squared, with a limiter for low velocity regimes to avoid overcorrection. This approach eliminates the need for background pressure, directly suppressing tensile instability in multi-resolution flows (Wang et al., 2024).

4. Advanced Techniques in Turbulent and Nonlinear Flows

• Speed-direction factorization in turbulence:

The velocity field is decomposed as $\bfu = u \bfr$ with $u = |\bfu|$ (speed) and $\bfr = \bfu / |\bfu|$ (direction), yielding coupled scalar and vector equations. This separation facilitates a scalar-based kinetic energy budget and reduces closure complexity in Reynolds-averaged turbulence modeling (Olshanskii, 2020).

• Lagrangian tensor velocity gradient modeling:

Statistical models for the evolution of the velocity gradient tensor at Kolmogorov scale leverage Lagrangian attention tensor networks (LATN), which parameterize closures (e.g., pressure Hessian and viscous Laplacian) as functions of temporal convolution features and classical invariants, yielding state-of-the-art accuracy and interpretability (Hyett et al., 10 Feb 2025).

5. Parameterization in Discretization, Registration, and Optimization Frameworks

• Band-limited representations in diffeomorphic mapping:

In PDE-constrained Large Deformation Diffeomorphic Metric Mapping (LDDMM), velocity fields are parameterized in a truncated Fourier basis (band-limited), reducing the dimension of the optimization problem and allowing momentum-conservation PDEs to be enforced as hard constraints during registration (Hernandez, 2018).

• Moving mesh and variational mesh velocity:

The mesh velocity in Arbitrary Lagrangian-Eulerian (ALE) and virtual element methods (VEM) is parameterized by a curl-free potential computed through variational forms, with solution transfer and node-insertion approaches designed to preserve local mass and global accuracy under mesh adaptation (Wells et al., 2022).

• Diffeomorphic deformation via scalar invariants:

Parameterization of deformation in registration can be achieved directly through the determinant of the Jacobian (volume change) and the curl (rotation) of an end-velocity field, constructing transformations that are inherently regularizing and diffeomorphic without explicit penalties (Sheikhjafari et al., 2022).

• Time-optimal path parameterization for robotics:

In the TOPP-DWR algorithm, robot motion along a B-spline path is parameterized via a discretized velocity profile subject to linear and conic constraints (on velocity, acceleration, angular rate, wheel rates). Slack variables transform the time-optimal problem into a convex second-order cone program, enabling efficient solution under practical robotic constraints (Li et al., 17 Nov 2025).

• Quaternion-based, nonsingular translational velocity parameterization:

Translational velocity is encoded as a pair $(v, q)$, with $v$ as the speed and $q$ a direction quaternion. This removes trigonometric singularities found in spherical coordinate representations, yielding globally nonsingular, numerically stable ODEs for both simulation and guidance applications (Miller et al., 2020).

6. Impact, Validation, and Comparative Performance

Empirical and theoretical validation of velocity parameterizations is a recurring theme:

• CNN-based ocean momentum closures yield RMSE reductions of $\sim60\%$ and bias reductions of $\sim60\%$ in subgrid eddy representation compared to unparametrized runs (Wang et al., 2024).
• Machine learning surrogates for air-entrainment velocity achieve $R = 0.999$ and RMSE $0.11\,\text{cm h}^{-1}$ globally, compared to $>1\,\text{cm h}^{-1}$ for bulk schemes (Zhou et al., 3 Feb 2026).
• Polynomial velocity laws in stellar wind models reduce mean velocity fit error from $\sim5\%$ (classical beta-law) to $1\%$ using three-term expansions, with similar gains in gradient accuracy (Krticka et al., 2011).
• In rarefied gas dynamics, adaptive velocity parameterization reduces CPU time and velocity-space size by factors of $10$–$40$ with negligible loss of solution accuracy (Chen et al., 2022).
• Lagrangian attention tensor network closures for turbulence yield 20–40% reductions in $L_2$ loss and improved phase-plane statistics relative to classical tensor basis neural networks (Hyett et al., 10 Feb 2025).

7. Limitations, Open Challenges, and Future Directions

Despite major advances, several open issues persist:

• The effectiveness of velocity parameterizations depends on the choice and expressivity of the basis or surrogate model, the quality and regime coverage of training data (when data-driven), and the enforcement of physical constraints, especially conservation laws.
• In adaptive and subdivided velocity spaces, threshold selection and conservation correction are critical and often chosen empirically (Chen et al., 2022).
• Neural closures for velocity or momentum require careful interpretability and validation against independent datasets and physical invariants (Wang et al., 2024, Hyett et al., 10 Feb 2025).
• Some parameterizations, particularly those based on spectral/modal expansion, may smooth out sharp local features or artifacts if insufficient modes are retained (Hernandez, 2018).
• The design and implementation of parameterizations for extreme regimes (superluminal, non-equilibrium, strong inhomogeneity) must incorporate rigorous self-consistency tests such as real-$\omega$ vs. real-$k$ equivalence in dispersive media (Tasgin, 2012).
• The development of modular, interpretable, and scale-aware parameterizations is a subject of ongoing research in multi-physics and multi-scale modeling.

In summary, velocity parameterization is a central, cross-disciplinary concept with diverse mathematical, physical, and computational incarnations. Its rigorous formulation and validation are indispensable for the accurate, efficient, and interpretable modeling of complex dynamical systems across physical, engineering, and computational domains.