PhysDial: Adapting Pretrained Video Models for PDE Simulation under Changing Physical Conditions

Yilong Dai, Yiming Sun, Shengyu Chen, Jiaqi Gong, Xiaowei Jia, Runlong YuarXiv 2026 · arXiv

TL;DR

PhysDial adapts large-scale pretrained video diffusion models to simulate complex physical systems (PDEs) under dynamic, changing conditions. By introducing "switch training"—training on trajectories where parameters change mid-stream—it overcomes the common failure of learned simulators to ignore their control inputs, effectively halving errors when forcing or coefficients shift during a simulation.

What Problem Does This Paper Address?

Simulating how a physical system—like an ocean current or a plasma—responds to changing external forces is critical for engineering design and control. While deep learning-based PDE solvers are significantly faster than traditional numerical methods, they suffer from a subtle but critical flaw: shortcut learning.

Most existing PDE foundation models are trained on trajectories where physical conditions (like viscosity or forcing amplitude) remain fixed. Because the physics of a system is often embedded in its recent history, a model can learn to "cheat" by inferring the conditions from past states rather than actually listening to its condition inputs. Consequently, if you change a parameter mid-simulation (e.g., turning up the wind speed), these models often fail to respond, continuing along the original path as if nothing happened.

Figure 1. When the forcing of a flow changes, the system leaves the future it was heading for. A simulator trained only on fixed-condition trajectories follows the change only in part. With switch training, PhysDial follows it.

Methods

The authors introduce PhysDial, which leverages the generative power of a pretrained video foundation model (Wan2.1, a 1.3B parameter diffusion transformer) and adapts it for autoregressive PDE simulation.

Key technical innovations include:

  • Condition Mapping: Instead of inventing new architectures, PhysDial routes physical inputs through the video model's existing conditioning pathways. Scalar coefficients modulate blocks via the noise-level pathway; spatial condition fields (like forcing patterns) enter as reference frames; and equation descriptions are treated as text prompts.
  • Switch Training: To break shortcut learning, the authors include a small fraction (~5%) of "switch trajectories" in the training data, where a condition changes abruptly at a random frame. This forces the model to rely on the condition input to determine which "branch" of the future to follow.
  • Anchor Trajectories: To improve extrapolation to unseen parameter values, they add "anchors"—trajectories simulated at a few specific values outside the standard training range.
  • Physical Latents: Physical fields are normalized and encoded into the latent space of a frozen video autoencoder, allowing the transformer to work with compressed representations while maintaining high-frequency details through a fine-tuned decoder.
Figure 2. Overall framework of PhysDial. Top. In each round, the same network predicts four future states from two past states and the current conditions. The last two predicted latents feed the next round, where the conditions can change. Bottom. Coefficients enter with the noise level, condition fields as reference frames, and equation text as a prompt. Snowflakes mark frozen modules; flames mark trained ones.

Datasets

The model was trained and evaluated on eight two-dimensional PDE families:

  1. The Well Benchmark: Six families including Gray-Scott (pattern formation), Rayleigh-Bénard (convection), and Euler equations (compressible flow).
  2. New Turbulent Simulations: Two newly simulated flows—Kolmogorov flow and Decaying turbulence—generated using a pseudo-spectral solver to provide rigorous ground truth for condition changes.

Key Insights and Ideas

  • The "Necessity" Hypothesis: A learned simulator uses a condition input only as far as its training data make that condition necessary for prediction. If the history always reveals the condition, the model ignores the "dial."
  • Data over Architecture: Adding a condition input to a pretrained model (like Walrus) after the fact doesn't work if the model has already learned to ignore it. The "fix" must come from the training data itself (switch training).
  • Video Models as World Simulators: Pretraining on massive video datasets provides a superior starting point for physics. Even without specific physics pretraining, PhysDial's adapted video backbone matched or exceeded dedicated PDE foundation models in turbulent flow accuracy.
  • Physical Fidelity: PhysDial maintains much better energy spectra at small scales compared to other foundation models. At step 64 in Kolmogorov flow, PhysDial had 1.07x the solver's small-scale energy, while the baseline Walrus model had 7.1x (indicating significant unphysical noise).

Key Results

  • Halving Error: On Kolmogorov flow, switch training halved the error following a change in forcing wavenumber or amplitude compared to models trained only on fixed conditions.
  • Response Rate: In counterfactual tests, PhysDial moved 75-95% of the way toward the "correct" new future after a change, whereas baseline models often moved less than 10%.
  • Extrapolation: Anchor trajectories reduced the cost of extrapolating to new forcing amplitudes by more than half.
  • Zero-Shot Physics: Despite no prior exposure to fluid dynamics before fine-tuning, PhysDial outperformed specialized PDE models (DPOT-H) on 13 out of 16 benchmark splits.
Figure 3. Tracking error under repeated changes (64-step Kol- mogorov rollouts). The wavenumber changes and returns to its initial value; the mixed sequence changes several conditions in turn. Changes occur every 16 steps except amplitude (32 steps). PhysDial has two seeds and Walrus one run.

Limitations and Future Work

  • Dimensionality: The current work is limited to 2D PDEs on regular grids. Extending this to 3D and irregular geometries is a natural next step.
  • Abrupt vs. Continuous: The study focused on abrupt "step" changes in conditions. Real-world systems often experience gradual, continuous changes (ramps), which require further investigation.
  • Autoencoder Bottleneck: The accuracy of a single step is currently limited by the reconstruction error of the frozen video autoencoder.
  • Scaling: While the 1.3B parameter model performed well, the authors noted that simply increasing the backbone size (e.g., to 14B) did not yield immediate gains at matched compute, suggesting that data quality and training strategy are currently more important than raw parameter count.

Summary updated

Abstract

Simulating how physical systems respond to changing conditions is central to control, design, and scenario studies. To our knowledge, prior PDE foundation models have not been trained or evaluated on changes within trajectories. Our counterfactual test measures how far predictions move toward a changed solver branch from the same state. Models trained on fixed conditions follow changes only partly. 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. Coefficients enter with the noise level, condition fields as reference frames, and equation text as a prompt. In Kolmogorov flow, switches within trajectories account for about 5% of the training samples. At matched compute, they halve the error after a change in forcing wavenumber or amplitude. A few trajectories at new amplitudes more than halve extrapolation error. Without physics pretraining, PhysDial matches or exceeds PDE foundation models under training conditions on two turbulent flows. At step 64 on the Kolmogorov test split, PhysDial has 1.07 times the solver's small-scale energy and Walrus 7.1 times. Our experiments suggest that a learned simulator uses a condition only as far as its training data make that condition necessary for prediction. Learning from trajectories that diverge under intervention can help scientific simulators predict what would happen if a system's forcing or parameters changed.
BibTeX
@misc{dai2026physdial,
  author = {Yilong Dai and Yiming Sun and Shengyu Chen and Jiaqi Gong and Xiaowei Jia and Runlong Yu},
  title = {{PhysDial: Adapting Pretrained Video Models for PDE Simulation under Changing Physical Conditions}},
  year = {2026},
  eprint = {2611.23498},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.23498},
}
Figure