PDE Surrogate

A numerical solver follows physical equations to advance a fluid or another evolving field. A learned surrogate tries to infer that evolution from examples. If accurate enough, it could make repeated simulations and parameter searches much faster.

Here one model sees four initial frames and predicts what comes next across sixteen equation families and three additional datasets. The systems include flowing fluids, active particles and layers of gas that mix while cooling. Their fields, boundaries and grids differ. A smooth-looking prediction can still erase a vortex or miss the mixing that scientists care about.

Two real physical systems, one required interface

Public upstream visualizations from The Well. These animations illustrate simulated physics; they are not this track’s held-out inputs or predictions. External images require access to the source host.

Visualizations: The Well / Polymathic AI, served by Flatiron Institute. Dataset license: CC BY 4.0. Each animation uses its source’s own scale; colors are not comparable between systems.

Follow one forecast through the model

01 · Observe

4 × C × X × Y

Four observed frames, with information about the domain and its boundaries.

02 · Normalize

Training statistics

The reference z-scores its fields. No statistics from future target frames.

03 · Predict

ConvNeXt U-Net

Combine broad spatial patterns with fine detail.

04 · Feed back

Advance context

Return to physical units; use the predicted fields to continue without future truth.

This is the PDEArena reference. It learns from compatible three-channel fields and repeats the last frame for other channel counts; a fully shared model remains the goal.

Why four group scores matter

Move one group’s mean error while holding the others fixed. This is a scoring calculator with illustrative values, not an experiment. The corpus contains many datasets but receives the same final weight as each transfer group.

Corpus VRMSE PDEArena VRMSE Active matter VRMSE Radiative layer VRMSE

Group rewards (0.667 + 0.667 + 0.333 + 0.333) ÷ 4 = 0.5000

R(v) = clip((1 − log₁₀(max(v, 0.01))) / 3, 0, 1). Improving any one group from 1 to 0.1 raises total reward by 1/12 ≈ 0.0833, regardless of its trajectory count. Changing only the corpus cannot erase failures in the other three groups.

For a simple fluid, the state may include a scalar field and two velocity components. Active matter also tracks particle orientation, producing eleven scored channels. The radiative layer includes density, pressure and velocity on a rectangular grid. These are physical quantities, not image color channels.

Trajectories are separated into training, validation and evaluation sets. Models start from random weights and see only initial frames and metadata during evaluation. The additional datasets are part of the declared learning setup, so “transfer” does not automatically mean predicting a wholly unseen kind of physics.

The main error is variance-normalized root mean squared error, or VRMSE. It compares prediction error with how much the true field varies across space. This allows fields with different units and amplitudes to be compared. A nearly uniform field makes even a small absolute mistake important relative to its remaining structure.

The equation corpus and each of the three additional datasets contribute equally after a logarithmic score mapping. Group errors of 1 and 0.1 give rewards of about 0.33 and 0.67. Reducing error by a factor of ten matters wherever it occurs; strong performance on one group cannot completely hide failure on another.

The reference uses an encoder–decoder network: compress spatial patterns, combine information at different scales, then reconstruct the next field. Connections between corresponding scales help retain fine detail. The inputs are normalized using training data and outputs return to physical units.

To continue a trajectory, the model must use its own predicted fields as context. This creates a mismatch with training examples that always start from the truth. A tiny phase error can shift a vortex, alter the next prediction and grow over time. The interactive example lets you compare this accumulation with repeatedly restarting from the true field.

The reference is trained on PDEArena fluid data. It uses its learned network on compatible three-channel inputs and otherwise repeats the last observed field, a simple strategy called persistence. Its overall reward therefore does not establish a universal solution.

Longer training reduces PDEArena error from about 0.297 to 0.282, while overall reward moves only from 0.345 to 0.347. The small overall gain makes sense: the other groups largely retain the same behavior. Repeating this experiment with multiple training runs would help establish how reliable the change is.

The reference’s error is about 0.30 on its training domain, compared with 1.52 on active matter and 1.36 on the radiative layer, where it uses persistence. This breakdown identifies a concrete gap: learning to represent different physical fields, rather than improving an already specialized fluid predictor alone.

Plotting error over time can reveal two different problems. A model that begins accurately but drifts may need training on longer self-generated trajectories. A model that fails immediately on orientation or pressure fields may need a better shared representation. Testing those changes separately, with the same data and computation, would make an improvement easier to explain.