pith. sign in

arxiv: 2602.17477 · v3 · submitted 2026-02-19 · 💻 cs.LG

Variational Grey-Box Dynamics Matching

Gurjeet Sangra Singh , Frantzeska Lavda , Giangiacomo Mercatali , Alexandros Kalousis This is my paper

Pith reviewed 2026-05-15 20:54 UTC · model grok-4.3

classification 💻 cs.LG
keywords grey-box modelingflow matchingvariational inferencedynamical systemsphysics-informed learninggenerative modelsODE/PDE learning
0
0 comments X

The pith

Incomplete physics models integrate into flow matching via variational latents to learn dynamics from observations while retaining interpretability.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a grey-box method that embeds incomplete ODE or PDE physics models directly into flow matching generative models. It learns system dynamics solely from observational trajectories by using a structured variational distribution with two latent encodings: one capturing missing stochasticity and multi-modal velocity, and the other encoding physics parameters under a physics-informed prior. Training proceeds simulation-free, avoiding the scalability issues of Neural ODEs, and an adaptation handles second-order dynamics. The approach targets problems where physics knowledge is partial yet still useful for guidance. Experiments on representative ODE/PDE cases and real-world weather forecasting show performance on par with or better than fully data-driven methods and prior grey-box baselines.

Core claim

A structured variational distribution is placed inside the flow matching framework with two latent encodings—one modelling the missing stochasticity and multi-modal velocity fields, the other treating physics parameters as a latent variable equipped with a physics-informed prior—allowing the model to learn dynamics from trajectories alone without ground-truth parameter values or numerical simulation.

What carries the argument

Structured variational distribution in flow matching with two latent encodings for stochasticity and physics parameters.

If this is right

  • The method achieves performance comparable to or exceeding fully data-driven flow matching and diffusion models on dynamical systems tasks.
  • It outperforms earlier grey-box baselines while keeping the physics model interpretable.
  • The framework extends naturally to second-order dynamics without additional simulation steps.
  • It applies directly to real-world forecasting problems such as weather prediction using observational data only.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Hybrid models of this form could reduce reliance on full numerical simulators during training for scientific applications.
  • The latent physics encoding may allow systematic diagnosis of which terms are missing from an initial model.
  • Extensions to stochastic partial differential equations or multi-scale systems become feasible once the variational structure is in place.

Load-bearing premise

The physics-informed prior on the latent physics parameters remains sufficiently informative and correctly specified even when the underlying ODE or PDE model is incomplete.

What would settle it

Controlled tests on known ODE systems where the supplied physics model deliberately omits a dominant term and the recovered trajectories diverge from ground truth or lose parameter interpretability.

read the original abstract

Deep generative models such as flow matching and diffusion models have shown great potential in learning complex distributions and dynamical systems, but often act as black-boxes, neglecting underlying physics. In contrast, physics-based simulation models described by ODEs/PDEs remain interpretable, but may have missing or unknown terms, unable to fully describe real-world observations. We bridge this gap with a novel grey-box method that integrates incomplete physics models directly into generative models. Our approach learns dynamics from observational trajectories alone, without ground-truth physics parameters, in a simulation-free manner that avoids scalability and stability issues of Neural ODEs. The core of our method lies in modelling a structured variational distribution within the flow matching framework, by using two latent encodings: one to model the missing stochasticity and multi-modal velocity, and a second to encode physics parameters as a latent variable with a physics-informed prior. Furthermore, we present an adaptation of the framework to handle second-order dynamics. Our experiments on representative ODE/PDE problems and real-world weather forecasting demonstrate that our method performs on par with or superior to fully data-driven approaches and previous grey-box baselines, while preserving the interpretability of the physics model. Our code is available at https://github.com/DMML-Geneva/VGB-DM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper presents Variational Grey-Box Dynamics Matching (VGB-DM), a method that integrates incomplete physics models (ODEs/PDEs) into flow-matching generative models for learning dynamics from observational trajectories alone. It employs a structured variational distribution with two latent encodings—one capturing missing stochasticity and multi-modal velocity, the other encoding physics parameters under a physics-informed prior—while avoiding simulation and Neural ODE issues; an adaptation for second-order dynamics is included. Experiments on synthetic ODE/PDE benchmarks and real-world weather forecasting claim performance on par with or better than fully data-driven and prior grey-box baselines, with preserved interpretability of the physics component.

Significance. If the central claims hold, the work provides a scalable, simulation-free route to hybrid grey-box modeling that combines the flexibility of flow matching with partial physical knowledge. This could meaningfully advance scientific machine learning applications where complete physics is unavailable, such as weather or other complex dynamical systems, while the open code release supports reproducibility.

major comments (2)
  1. [Abstract] Abstract: the central claim that the method 'preserves the interpretability of the physics model' rests on the second latent encoding recovering meaningful parameters under model incompleteness, yet no explicit form of the physics-informed prior, no parameterization details, and no recovery diagnostics (e.g., correlation with ground-truth quantities or robustness under misspecification) are supplied; this is load-bearing for the grey-box novelty.
  2. [Method] Method description: the structured variational distribution inside the flow-matching objective is introduced at a high level, but the interaction between the stochasticity latent and the physics-parameter latent (including whether the variational family permits trade-off that could absorb residuals) is not formalized with equations; without this, it is impossible to assess whether performance parity can occur without true interpretability.
minor comments (2)
  1. [Experiments] Experiments: error bars, multiple random seeds, or statistical tests should be reported for the weather-forecasting results to substantiate the 'on par or superior' claim.
  2. [Abstract] The simulation-free training and code availability are positive for reproducibility and should be highlighted more explicitly in the abstract.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the potential of VGB-DM. We agree that the interpretability claim and the formalization of the variational structure are central and currently underspecified. Below we address each major comment and outline the revisions we will make.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the method 'preserves the interpretability of the physics model' rests on the second latent encoding recovering meaningful parameters under model incompleteness, yet no explicit form of the physics-informed prior, no parameterization details, and no recovery diagnostics (e.g., correlation with ground-truth quantities or robustness under misspecification) are supplied; this is load-bearing for the grey-box novelty.

    Authors: We agree that the current manuscript does not supply sufficient detail to substantiate the interpretability claim. In the revised version we will (i) explicitly state the physics-informed prior (a Gaussian centered on nominal parameter values with variance scaled by expected model incompleteness), (ii) give the exact parameterization of the second latent encoder (MLP outputting mean and diagonal covariance of the variational posterior over physics parameters), and (iii) add recovery diagnostics: Pearson correlations between inferred and ground-truth parameters on synthetic benchmarks, plus robustness plots under controlled misspecification of the nominal values. These additions will be placed in a new subsection of the method and supported by new figures in the experiments. revision: yes

  2. Referee: [Method] Method description: the structured variational distribution inside the flow-matching objective is introduced at a high level, but the interaction between the stochasticity latent and the physics-parameter latent (including whether the variational family permits trade-off that could absorb residuals) is not formalized with equations; without this, it is impossible to assess whether performance parity can occur without true interpretability.

    Authors: We accept that the interaction between the two latents must be formalized. In the revision we will insert the full variational objective with the joint posterior q(z_s, z_p | trajectory) = q(z_s | trajectory) q(z_p | trajectory, z_s), the reparameterized sampling, and the precise decomposition of the flow-matching loss into a physics-informed term (conditioned on z_p) and a residual term (handled by z_s). We will also add a short analysis showing that the physics-informed prior on z_p, together with the conditional dependence structure, prevents z_s from fully absorbing systematic physics residuals. These equations and the accompanying discussion will appear in Section 3. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method is an architectural extension of flow matching with external physics-informed prior

full rationale

The derivation introduces a structured variational distribution inside the existing flow-matching framework, using one latent for stochasticity and a second for physics parameters equipped with a physics-informed prior. This prior and the overall grey-box construction are modeling choices imported from outside the paper rather than fitted quantities or self-referential definitions. No equations reduce by construction to the target performance metric, no load-bearing self-citations appear, and the central claims rest on empirical comparisons to data-driven baselines and prior grey-box methods. The approach is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the existence of a usable physics-informed prior for the latent parameters and on the assumption that the incomplete ODE/PDE model still provides a useful inductive bias even when terms are missing.

free parameters (1)
  • latent dimension for physics parameters
    Dimensionality of the second latent encoding is chosen by the authors and not derived.
axioms (1)
  • domain assumption The incomplete physics model remains a valid structural prior even when terms are missing.
    Invoked when the method is described as integrating incomplete physics models directly into the generative process.

pith-pipeline@v0.9.0 · 5530 in / 1241 out tokens · 32066 ms · 2026-05-15T20:54:42.232696+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.