← News

Models & toolsSources: arXiv · 19 Nov 2026

PhysDial adapts video diffusion models to simulate PDEs under changing conditions

A new arXiv paper introduces PhysDial, a method that tunes a pretrained video diffusion transformer so its PDE simulations respond to changed physical conditions mid-trajectory.

In plain words

Partial differential equations, or PDEs, are equations that describe how physical systems like fluid flows evolve over time. The authors say existing PDE foundation models were not trained on conditions changing within a single simulation, so they barely react when conditions shift. PhysDial adapts a pretrained video model to handle such shifts, roughly halving prediction errors after a change in their tests.

What was announced

The paper, posted to arXiv by the PhysDial authors, reports that prior PDE foundation models have not, as far as the authors know, been trained or evaluated on changes occurring within trajectories. The team proposes a counterfactual test that measures how far a model's predictions move toward a changed solver branch from the same starting state. Under this test, models trained on fixed conditions follow changes only partly, and a PDE foundation model given a condition input after pretraining barely responds.

PhysDial adapts a pretrained video diffusion transformer to eight two-dimensional PDE families, feeding in coefficients alongside the noise level, condition fields as reference frames, and equation text as a prompt. In Kolmogorov flow, where within-trajectory switches make up about 5% of training samples, the authors report halved error after a change in forcing wavenumber or amplitude at matched compute, and a few trajectories at new amplitudes more than halved extrapolation error. Without physics pretraining, PhysDial matched or exceeded PDE foundation models on two turbulent flows under training conditions; at step 64 on the Kolmogorov test split it showed 1.07 times the solver's small-scale energy versus 7.1 times for Walrus. The authors suggest a learned simulator uses a condition only as far as its training data make it necessary.

What it does not show

  • The evidence is the paper's abstract only; the full article's methods, experiments and peer review status are not covered, and the reported results are the authors' own and not independently confirmed.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.