Teaching a neural operator raised on clean simulations to forecast messy, real water-tunnel flow.
-64748b?style=flat-square){kind=icon}
Method · Results · Experiment ledger · Reproducing
An illustration I generated, not competition data: the wake of a NACA4418 airfoil from a small lattice-Boltzmann simulation (left), and the same flow degraded the way PIV measurements are (right). How it was made
The organizers' pretrained models learned from clean simulations like the left panel. They're graded on measurements from a real water tunnel, taken with particle image velocimetry (PIV), which look a lot more like the right: grainy, blurred, and never quite in step with the simulation. (The competition website shows the real side-by-side.)
Track 1 of the RealPDE Competition asks a simple-sounding question: can a model trained mostly on the left panel make useful predictions about the right one?
The task: watch 20 frames of measured velocity (a 32 × 64 grid, u and v), then predict the next 20.
Pixel accuracy is only part of the score. Forecasts are also judged on turbulent kinetic energy (TKE), the mean wake profile (MVPE), inference speed, and whether the uncertainty bounds are honest (SPS).
I didn't invent a new architecture. I started from the organizers' simulation-pretrained convolutional neural operator (CNO) and spent the competition working out how to adapt, correct and trust it on real data. The recipe behind my best submission, E037:
| # | Step | Why it helped |
|---|---|---|
| 1 | Longer fine-tuning on real flow. 3000 updates (up from 600), starting from the simulation checkpoint. | The longer-budget recipe improved all four quality subscores on two separate held-out regime splits. |
| 2 | A scale-balanced loss. Each window's error is divided by that window's own flow energy. | Stops fast, high-energy flows from drowning out everything else. |
| 3 | One small correction network. Keep the main model fixed and train a lightweight adapter, including practice forecasting further ahead. | Cheap, stable refinement. One adapter held up where three didn't (see below). |
| 4 | Persistence for the first four frames. Copy the last observed frame forward. | The next few frames look a lot like the present. Persistence is a strong short-horizon baseline. |
| 5 | Calibrated uncertainty bounds. Pair each forecast with a range of plausible values. | Makes the model say how unsure it is, which the SPS metric rewards. |
| 6 | Independent samples. No batching tricks, caching or peeking across evaluation windows. | Each prediction is the same no matter what it's evaluated alongside. |
The held-out evidence came from separate experiments, E034 and E038. E037's backbone was fitted on all 81 valid real trajectories, with no clean public holdout. The full recipe, technical details and caveats are in METHOD.md.
The project spans records E001–E043 across August and September 2026, including planned and paused experiments. E037 delivered the best result in August; later work did not surpass it. Below are selected official development-leaderboard scores I recorded from Codabench.
xychart-beta
title "Official development-leaderboard score (final_score)"
x-axis [E008, E013, E021, E029, E037]
y-axis "final_score" 74 --> 78
line [74.84, 75.57, 75.87, 75.88, 77.31]
| Submission | What changed | Score |
|---|---|---|
| E008 | First package whose predictions don't depend on batch composition | 74.84 |
| E013 | Persistence for the first forecast frames | 75.57 |
| E021 | A residual adapter on top of the CNO backbone | 75.87 |
| E029 | All-data training with a rollout-trained adapter | 75.88 |
| E037 | Long-budget backbone (3000 updates) + single robust adapter | 77.31 |
E037's full score breakdown
| What was measured | Metric | Score |
|---|---|---|
| Flow-field accuracy | Rel-L2 | 94.023891 |
| Turbulent energy | TKE | 73.827351 |
| Mean wake profile | MVPE | 93.166385 |
| Inference speed | Time | 85.988280 |
| Uncertainty quality | SPS | 32.361125 |
| Overall | final_score | 77.307902 |
E037 improved every metric over E029. Scores are the development-leaderboard results I recorded from Codabench. They aren't a final private-test ranking or a fresh reproduction. Full results and caveats →
Honestly, the failures taught me as much as the wins, so they stay in the record.
- 💥 The ensemble that fell off a cliff. A three-adapter ensemble passed every local gate I had, then scored 7.96 on the server. Stress tests later made it blow up on rare extreme inputs, and a review found a separate wrapper bug, so the exact cause was never pinned down. The lesson: robustness beats cleverness, and E037 went back to a single adapter.
- 🚫 A local pass is not a server pass. My final attempt, E043, passed local container checks and then failed evaluation. I never got the log, so the cause is unknown. It doesn't replace E037.
- 📉 Wavelet losses, FiLM conditioning, heteroskedastic heads, spatial uncertainty maps. Each got its own experiment and some helped locally, but none made it into the final recipe. The simpler path kept winning.
- 🔁 Validate on whole regimes. Holding out entire Reynolds-number / angle-of-attack conditions, not random windows, was what kept my local numbers honest. See VALIDATION.md.
Every experiment, the dead ends included, is in the experiment ledger.
| If you want to… | Go to |
|---|---|
| Understand the approach | METHOD.md |
| See the numbers and their caveats | RESULTS.md · VALIDATION.md |
| Follow the whole research story | LEDGER.md · literature notes · task reference |
| Read the code | src/realpde_t1/ · scripts/ · configs/ · submission/ |
| Rebuild the environment | REPRODUCING.md · EXTERNAL_DEPENDENCIES.md · environment-locks/ |
This is a research archive, not a pretrained model you can download and run.
- ✅ Included: my source code, configurations, regression tests, the experiment ledger and result notes, and exact pinned environments.
- ❌ Not included: competition data, trained weights, generated submission archives, or the organizer starter kit. The data and checkpoints were deleted when the competition ended.
⚠️ Not re-run: I haven't done a fresh end-to-end reproduction since the cleanup. Remaining reconstruction gaps include E037 checkpoint assembly, missing intermediate evaluation files, and unverified access to the historical starter kit. See REPRODUCING.md.
You can check the archive right away. No GPU, data or packages needed:
python3 -B scripts/check_archive.py
python3 -B scripts/publication_audit.pyThese check source integrity and release hygiene, not model quality. To go
further, follow REPRODUCING.md. Clone as RealPDE_T1 if you
want paths to match the original workspace layout.
Sibling project:
realpde-track2covers Track 2, where a model forecasts an ongoing real-flow stream and adapts as each new measurement arrives.
Built by Pranay Vandanapu, with AI coding assistance (details). Huge thanks to the RealPDE organizers and the RealPDEBench team for the dataset, the baselines and a genuinely interesting problem. The flow animation is my own illustration, made with a small script (how it was made).
My original code is MIT-licensed. The model/data stack it depends on is not: RealPDEBench is CC BY-NC 4.0, and some upstream utilities carry research-only terms, so please don't treat this repository as uniformly MIT. Details: LICENSING.md · THIRD_PARTY_NOTICES.md · PUBLICATION_AUDIT.md.