Zero-Shot Physics-Consistent Sampling

• Zero-Shot Physics-Consistent Sampling is a methodology that generates samples conforming to physical laws using pretrained generative models without retraining.
• It employs techniques like hard constraint projection, proximal guidance, and flow-matching to ensure models satisfy constraints in applications such as PDE modeling and climate downscaling.
• Empirical results show that these approaches achieve significantly lower constraint residuals and higher generative fidelity compared to traditional penalty-based retraining methods.

Zero-shot physics-consistent sampling refers to a class of generative inference algorithms that, given a pretrained model and a set of physical constraints (e.g., specified by partial differential equations, boundary conditions, or measurement operators), produce samples that are both statistically aligned with the model prior and strictly (or nearly) satisfy the physical constraints—without retraining or fine-tuning on the new tasks or measurements. This paradigm is of particular importance in computational physics, inverse problems, and scientific machine learning, where the cost of retraining on new physical domains or observables is prohibitive, and exact physical consistency with governing laws is mandatory.

1. Foundational Principles of Zero-Shot Physics-Consistent Sampling

Zero-shot physics-consistent sampling is grounded in two core requirements: (1) consistency with a learned data-driven generative prior, and (2) exact or approximate satisfaction of physical laws or observables at inference time. The "zero-shot" attribute emphasizes that adaptation to new constraints (e.g., novel PDEs, boundary conditions, or measurements) is performed entirely at inference, not during model training.

This concept has been formalized across multiple generative frameworks:

• Diffusion models: Approaches augment the reverse sampling chain by incorporating measurements (likelihood guidance) or physics-based regularization dynamically, while keeping the diffusion prior fixed (Tie et al., 29 Jan 2026, Alçalar et al., 2024).
• Flow-matching ODE models: Recent zero-shot algorithms dynamically correct the ODE trajectories during sampling, projecting states onto physics-manifolds or measurement-consistent sets without altering the trained prior (Utkarsh et al., 4 Jun 2025, Yu et al., 28 Jan 2026).

Physics consistency can be enforced as a hard constraint (e.g., via projection onto the physical manifold) or as a tight penalty (residual minimization); zero-shot protocols favor hard enforcement for scientific domains.

2. Algorithmic Methodologies

Several algorithmic templates have emerged:

Hard Constraint Projection

Physics-Constrained Flow Matching (PCFM) utilizes tangent-space projections and iterative Newton-type solvers to ensure the ODE sample trajectory terminates on the constraint manifold $g(u)=0$. At each step, Gauss–Newton updates are used to orthogonally project to $\{g(u)=0\}$, followed by backward interpolants and optional penalty minimization for nonlinear constraints. The continuous projection mechanism guarantees, under mild Jacobian regularity, that generated samples are physically admissible (Utkarsh et al., 4 Jun 2025).

Proximal Guidance

ProFlow adopts a two-step alternating scheme: (1) after an Euler prediction along the learned flow, a proximal optimization projects the state onto the intersection of the physics manifold and observation-consistent set; (2) a flow-consistent interpolation maps the resulting state back onto the learned flow path to maintain distributional fidelity. This two-phase alternation reconciles physical exactness with prior consistency, and admits a local MAP interpretation under a Bayesian framework (Yu et al., 28 Jan 2026).

Regularized Consistency Training

Consistency Training with Physical Constraints (CT-Physics) employs a two-stage process: first, a mapping from noise to data is learned via consistency training; then, a physical residual penalty is imposed to drive the generator output toward the physical solution manifold. At inference, sampling is performed in a single step, yielding approximate physics-consistent samples with minimal residual $\|\mathcal{R}(x_0)\|$ (Chang et al., 11 Feb 2025).

Zero-Shot Physics-Guided Diffusion Posterior Sampling

Diffusion-based frameworks, such as ZSSD and ZAPS, modify the reverse sampling chain by introducing likelihood or physics gradients that guide the denoising process toward measurement or PDE consistency. These approaches introduce learnable per-step guidance strengths or loss-weighting parameters via zero-shot tuning, while respecting the pretrained generative prior (Tie et al., 29 Jan 2026, Alçalar et al., 2024).

3. Mathematical Formalism and Theoretical Guarantees

Suppose the physical constraint is specified as $g(u)=0$, which could represent boundary conditions, conservation laws, or discretized PDE residuals.

• In PCFM, the sampling ODE is modified to

$$\frac{du}{d\tau} = v_\theta(u, \tau) - J(u)^\top [J(u)J(u)^\top]^{-1} J(u) v_\theta(u, \tau)$$

where $J(u) = \nabla g(u)$. Under twice differentiability and full-rank Jacobian, this projection ensures exact constraint satisfaction at convergence (Utkarsh et al., 4 Jun 2025). Theoretical results establish that final Newton–Schur projection returns a sample satisfying $\|g(u)\|$ up to machine precision.

• ProFlow solves, at each step, the subproblem

$$u_1^\star = \arg\min_u \left\{ \frac{1}{2\sigma_t^2}\|u - \hat{u}_1\|_2^2 + \frac{1}{2\sigma_{\rm obs}^2}\|\mathcal{H}[u] - y\|_2^2 \right\}\quad\text{subject to}\quad \mathcal{L}(u)=0$$

and retrojects onto the flow manifold via OT (optimal transport) interpolation. This decomposition ensures samples remain statistically plausible while exactly satisfying the physics manifold (Yu et al., 28 Jan 2026).

• Diffusion-based zero-shot sampling, as in ZSSD and ZAPS, leverages Tweedie’s formula and the posterior score decomposition

$$\nabla_{x_t}\log p(x_t\mid y) = \nabla_{x_t}\log p(x_t) + \nabla_{x_t}\log p(y\mid x_t)$$

where the latter term is approximated using adjoint physics or measurement models, and the guidance strength is dynamically tuned for measurement/physical consistency (Tie et al., 29 Jan 2026, Alçalar et al., 2024).

4. Application Domains and Empirical Performance

Zero-shot physics-consistent samplers have been applied to a range of challenging inverse and forward problems:

• PDE-Constrained Generative Modeling: ProFlow and PCFM have demonstrated strict enforcement of boundary conditions and conservation properties on Poisson, Helmholtz, Darcy, and viscous Burgers’ equations, achieving orders-of-magnitude lower residuals than soft-penalty or unconstrained approaches, as well as improved MMSE, SMSE, and Fréchet Poseidon Distance metrics (Utkarsh et al., 4 Jun 2025, Yu et al., 28 Jan 2026).
• Climate Downscaling: ZSSD enables zero-shot statistical downscaling from coarse GCM fields to high-resolution climate fields, preserving extreme event statistics (e.g., 99th percentile errors) without exposure to paired training data (Tie et al., 29 Jan 2026).
• Computational Imaging and Phase Retrieval: GedankenNet and Phy-ZSN frameworks enforce physics constraints via forward model cycles, enabling accurate hologram reconstruction and phase retrieval without paired data or retraining, and robust to physical perturbations (Huang et al., 2022, Kumar, 2021).
• Physics-Based Video Generation: MotionCraft warps the latent space of diffusion models according to simulator-derived optical flow, producing temporally coherent, physically plausible video without supervised video training (Aira et al., 2024).

Application	Physics Consistency	Example Problems
PDE generative modeling	Hard constraints	Poisson, Helmholtz, Darcy, Burgers, Navier–Stokes
Climate downscaling	Likelihood/composite	GCM to ERA5, extremes (cyclones, rainfall)
Computational imaging	Physics cycle	Holography, inverse phase retrieval
Video generation	Prescribed motion	Rigid body, fluids, natural scenes

5. Comparison with Retraining and Penalty-Based Methods

Traditional strategies enforce physics by:

• Incorporating soft penalties (e.g., residual terms in the loss) during retraining of generative models.
• Designing architecture-specific biases that encode conservation properties.

These approaches are limited by susceptibility to residual drift, inability to guarantee strict satisfaction, and high computational cost for each new task or set of measurements. In contrast, zero-shot physics-consistent sampling:

• Operates directly on pretrained models, requiring no retraining or fine-tuning when the physical constraint or measurement setup changes (Yu et al., 28 Jan 2026, Utkarsh et al., 4 Jun 2025).
• Provides (for hard-constraint methods) theoretical guarantees of exact constraint satisfaction.
• Enables rapid adaptation to new inverse problems, measurement patterns, or physics domains.

Empirical results consistently show that hard-constraint zero-shot samplers (PCFM, ProFlow) outperform both unconstrained and penalty-based learned baselines on constraint satisfaction, generative accuracy, and distributional metrics (Utkarsh et al., 4 Jun 2025, Yu et al., 28 Jan 2026). For example, ProFlow's mean residual error on Poisson is 1.17×10⁻⁹ versus 8.79×10⁻⁹ for DiffusionPDE and 6.60×10⁻⁹ for ECI (Yu et al., 28 Jan 2026).

6. Limitations and Future Directions

While zero-shot physics-consistent approaches represent a significant advance, current limitations include:

• The requirement to solve a proximal or projection subproblem at each sampling step, which can be computationally intensive for high-dimensional fields (Yu et al., 28 Jan 2026).
• Difficulties with highly nonlinear, turbulent, or chaotic PDEs where gradient-based solvers may not converge globally or efficiently.
• Limited global convergence guarantees for the alternating proximal–flow schemes; analysis is currently local to the constraint manifold (Yu et al., 28 Jan 2026).
• For some frameworks (e.g., CT-Physics), approximate rather than exact satisfaction of nonlinear constraints may occur if the physics regularizer is not sufficiently weighted or if the optimization landscape is ill-conditioned (Chang et al., 11 Feb 2025).

Potential research targets include integration of learned proximal solvers, multiscale control for turbulent phenomena, scalable operator splitting techniques, and formalization of global convergence results.

7. Significance and Outlook

Zero-shot physics-consistent sampling frameworks are emerging as central tools for physical simulation, scientific inference, and general purpose inverse problems. By leveraging strong data-driven priors and enforcing physical admissibility without retraining, these methods offer scalability, adaptivity, and rigor that are infeasible for classical solvers or purely data-driven optimization. Applications range from climate modeling, PDE simulation, and computational imaging to video synthesis with prescribed physics. Ongoing advances in projection algorithms, flow-based architectures, and large-scale physical benchmarks continue to expand the reach and robustness of zero-shot physics-consistent generative sampling (Yu et al., 28 Jan 2026, Utkarsh et al., 4 Jun 2025, Tie et al., 29 Jan 2026, Chang et al., 11 Feb 2025, Huang et al., 2022, Alçalar et al., 2024, Kumar, 2021, Aira et al., 2024).