Neural Velocity Fields: Principles & Applications

• Neural Velocity Field is a continuous mapping that uses neural networks to convert spatiotemporal coordinates into velocity vectors while enforcing physical priors.
• It integrates physics-informed training methodologies, including divergence-free constraints and smoothness penalties, to achieve realistic flow representations.
• Applications span fluid simulation, video reconstruction, cosmological mapping, and medical imaging, demonstrating significant accuracy improvements over classical methods.

A neural velocity field is a continuous mapping from spatial (and sometimes temporal) coordinates to velocity vectors, represented or parameterized by neural networks. This concept encompasses diverse applications, including data-driven simulation, inverse problems in physics, video-based scene analysis, cosmological field reconstruction, and medical flow imaging. By leveraging neural networks—often multilayer perceptrons (MLPs) or convolutional architectures—with explicit regularization or physics-motivated inductive biases, neural velocity fields combine data fidelity with the ability to enforce physical or geometric priors, overcome discretization artifacts, and facilitate differentiable learning-based pipelines.

1. Foundational Definitions and Representational Forms

A neural velocity field can be formally defined as a function

$$\mathbf{v}_\theta: (\mathbf{x}, t) \mapsto \mathbb{R}^d$$

where $\mathbf{x} \in \Omega \subset \mathbb{R}^d$ and $t \in \mathbb{R}$, and where the neural network with parameters $\theta$ is trained to either approximate continuous ground-truth data or satisfy constraints from physics-informed objectives.

Several orthogonal axes of design occur:

• Architectures: MLPs with sinusoidal, ReLU, or tanh activations; convolutional UNets for gridded 3D data; hybrid models combining neural and analytical elements.
• Output modalities: Direct prediction of velocity components ($\mathbf{v}$), velocity potentials ($\phi$), or bases/coefficients for reduced (kinematic) velocity representations.
• Spatio-temporal parameterization: Functions of space alone ($\mathbf{v}(x)$) for stationary cases, or fully space-time ($\mathbf{v}(x, t)$) for dynamic settings.
• Physical or geometric priors: Embedded divergence-free, orthogonality, smoothness, boundary alignment, or direct PDE satisfaction.

For instance, in the context of video-based dynamic radiance fields, neural velocity fields are MLPs mapping ($\mathbf{x}, t$) to $\mathbb{R}^3$, used to advect NeRF features (Li et al., 2023). In mesh-free fluid simulation, the velocity field is expanded as a linear combination of neural basis fields parameterized by MLPs and global coefficients, enforcing crucial invariants (divergence-free, boundary, orthogonality, smoothness) (Liu et al., 22 Apr 2025). In MRI flow imaging, neural fields map spatial and temporal coordinates to parametrizations of phase-encoded complex images, from which velocity is then analytically derived (Arratia et al., 29 Sep 2025).

2. Network Training Methodologies and Physics-Informed Losses

Training neural velocity fields typically involves a hybrid loss composed of both observation/likelihood terms and physics-informed regularizers or constraints:

• Supervised/likelihood-based: Direct mean squared error (MSE) between network predictions and observed or simulated velocities; photometric/image losses where ground truth is observable only indirectly (e.g., rendered intensity via differentiable rendering pipelines).
• PDE and physics constraints: Soft or hard enforcement of governing equations (e.g., incompressibility, momentum conservation, advection transport, Navier–Stokes residuals, Stokes or Poisson equations).
• Basis regularization: For basis-expansion models, explicit losses for basis vector orthogonality, length normalization, divergence, and boundary alignment (Liu et al., 22 Apr 2025).
• Automatic differentiation: All spatial and temporal derivatives required by physical regularization terms or PDE constraints are computed via automatic differentiation of the neural field outputs.

Examples include:

• Physics-informed neural networks (PINNs) reconstructing the velocity field in pool-fire simulations by minimizing both data-fitting and residuals of the reacting Navier–Stokes system over collocation points (Sitte et al., 2022).
• Composite training objectives with multiple jointly optimized networks: in NVFi, separation of keyframe radiance field and interframe neural velocity field, trained by a combination of photometric losses and physics-informed PINN losses (divergence-free, Euler equation) (Li et al., 2023).
• Hybrid regimes: in HyFluid, the neural velocity field is split into a smooth neural base and a vortex-particle sum, with losses representing advection, incompressibility, and laminar regularization, plus rendering-based data terms (Yu et al., 2023).

Training typically leverages large synthetic datasets, Monte Carlo sampling for coverage of spatial domains, batch gradient descent, and staged/alternate optimization in multi-component architectures.

3. Applications: Simulation, Inverse Problems, and Scientific Inference

Neural velocity fields have been applied across a spectrum of scientific and engineering domains:

Domain	Neural Velocity Field Role	Reference
Computational Fluids	Basis for mesh-free velocity representation and interactive flow animation	(Liu et al., 22 Apr 2025)
Video-based Reconstruction	3D motion learning, future frame prediction, semantic decomposition	(Li et al., 2023, Yu et al., 2023)
Inverse Rheology	Joint learning of velocity field and constitutive law from flow data	(Lardy et al., 21 Jun 2025)
Astrophysics	Large-scale velocity field reconstruction from galaxy distributions	(Veena et al., 2022, Lilow et al., 2024)
Medical Imaging	Flow (blood velocity) recovery from undersampled MRI and MRV data	(Arratia et al., 29 Sep 2025, Sengupta et al., 2021)
Computer Vision	Deformation of neural implicit surfaces via explicit neural velocity field	(Sang et al., 23 Jan 2025)

In cosmological field recovery, neural fields outperform Wiener filters for reconstructing nonlinear peculiar velocity fields, capturing non-Gaussian features and bulk flows more faithfully than classical analytical priors (Veena et al., 2022, Lilow et al., 2024). For viscoplastic material inference, the neural streamfunction network, together with learnable constitutive-law parameters, enables direct discovery of hidden rheological laws solely from observed velocities (Lardy et al., 21 Jun 2025). In high-acceleration MRI reconstruction, neural fields yield continuous, temporally coherent representations of complex-valued images from which velocity can be robustly determined, even at extreme data undersampling (Arratia et al., 29 Sep 2025).

4. Imposed Physical Priors and Structure-Preserving Mechanisms

A central advantage of neural velocity fields in scientific settings is the ability to enforce domain-specific physical structure to ensure physically meaningful, regular, and interpretable velocity maps:

• Divergence-free constraint: Essential for modeling incompressible flows, enforced via explicit loss or built-in model structure (e.g., streamfunction parameterization).
• Boundary alignment: Tangency constraints at boundaries, essential in fluid simulation and meshing schemes (enforced via cosine similarity of velocity and surface normals penalized near boundaries) (Liu et al., 22 Apr 2025).
• Smoothness and regularity: Penalties on gradient norms, Jacobians, or higher derivatives suppress spurious oscillations and ensure spatial coherence.
• Orthogonality: In multi-basis expansions, mutual orthogonality prevents redundancy and ensures interpretability (Liu et al., 22 Apr 2025).
• Direct physics satisfaction: By predicting quantities such as velocity potentials or streamfunctions and computing velocities via differentiation, certain conservation laws are exactly satisfied (e.g., the linearized momentum equation in acoustic modeling via scalar potential fields) (Masuyama et al., 23 Mar 2026).
• Volume preservation and diffeomorphism: Penalized divergence and Hilbert-space norms restrain deformation fields to regular, non-intersecting transformations (Sang et al., 23 Jan 2025).

This principled enforcement of physical structure via differentiable neural surrogates allows these models both to robustly interpolate sparse, noisy, or partial data, and to generalize to domains unseen during training.

5. Quantitative Evaluation, Limitations, and Experimental Insights

Empirical benchmarks across domains demonstrate significant improvements from neural velocity fields compared to analytical or non-neural approaches:

• Factor-of-two reduction in velocity MSE over Wiener filter in cosmic-field reconstruction (Veena et al., 2022).
• Extrapolation PSNR improvements of 7–8 dB versus baseline models for 3D video motion prediction (Li et al., 2023).
• Up to 4% velocity error and preserved anatomy in accelerated MRI at 32× undersampling (Arratia et al., 29 Sep 2025).
• Divergence-free violations at L2 norm below $\mathbf{x} \in \Omega \subset \mathbb{R}^d$0 and visually precise recovery of physical features in fluid animators (Liu et al., 22 Apr 2025).
• Robustness to high noise in MRV and fire flow reconstruction (reconstruction errors below 8% at SNR=2.5–20 dB) (Sengupta et al., 2021, Sitte et al., 2022).
• PEG-based model-selection weights in PINNs achieving >90% correct rheology law identification under various synthetic flows (Lardy et al., 21 Jun 2025).

Key experimental findings include the criticality of "keyframe + transport" regime for dynamic scene learning (Li et al., 2023), the need for hybrid neural-vortex field decompositions to represent turbulence (Yu et al., 2023), and the positive regularization effect from implicit biases of continuous neural representations ("spectral bias") in ill-posed or data-scarce settings.

Limitations noted include dependence on representative training data for inverse or observational pipelines, sensitivity to incomplete physics enforcement (e.g., only momentum, not all governing PDEs), and, in shape-deformation fields, reduced fidelity for highly nonlocal or topologically complex deformations. However, embedded model selection and physics-based regularization generally mitigate overfitting, while advances in differentiable programming architectures support broad extensibility (e.g., to 3D or time-resolved problems).

6. Extensions and Future Directions

Recent and emerging trends in the field include:

• Multiphysics and multi-modal fields: Extending velocity field models to multi-material, multi-phase, or coupled PDE problems (e.g., Cahn-Hilliard–Navier–Stokes, elastodynamics).
• Adaptive neural architectures: Construction of spatially sparse, hierarchical, and multi-resolution neural buffers for scalable high-fidelity simulation (Deng et al., 2023).
• Hybrid analytical–neural decompositions: Combination of neural main-field and vortex-particle residual for turbulence representation and improved recovery of fine-scale flow phenomena (Yu et al., 2023).
• Model-agnostic inference: Simultaneous learning of the velocity field and identification of the most appropriate constitutive or rheological law (Lardy et al., 21 Jun 2025).
• Data assimilation and domain adaptation: Training on synthetic simulations then adapting to experimental data, or fine-tuning physics priors for better generalization to novel scenes (e.g., transfer of dynamic field models between unrelated geometries via motion transfer) (Li et al., 2023).
• Direct PDE-satisfying formulation via potentials: Predicting potentials (stream, velocity, or others) so that their derivatives by construction satisfy certain governing equations, thus removing the need for explicit penalization (Masuyama et al., 23 Mar 2026).

As neural velocity field frameworks become more systematic and physically robust, new applications are emerging in in situ simulation, experimental design, data assimilation, sensor fusion, and real-time model-predictive control, emphasizing the generality of continuous, differentiable, and constraint-preserving neural surrogates for velocity fields across science and engineering.