AI4Science-WestlakeU/RealPDE-Competition-Data

RealPDE Competition Data (NeurIPS 2026)

Training data and baseline checkpoints for the NeurIPS 2026 RealPDE Competition. This is a mirror of the competition's Google Drive release, hosted here because the Drive link runs into a per-file anonymous download quota when many people fetch it at once.

Both tracks share this release:

• —Track 1, Sim2Real — codabench.org/competitions/17363
• —Track 2, LTTTA — codabench.org/competitions/17385

Contents

train_sim.tar.gz              simulated training trajectories, for pretraining
train_real.tar.gz             real-world PIV training trajectories, for finetuning
example_data/3750_0.h5        one real trajectory, for checking shapes locally
baseline_checkpoints/
  pack_ckpt_fp16.py           complex-safe fp16 packing helper
  sim_pretrain/               CNO, FNO, FNO-fp16, Transolver, simulation only
  sim_real_ft/                the same four, finetuned on real data

The hidden evaluation set is not part of this release and is not published anywhere. It exists only inside the Codabench evaluation container.

`train_real/7575_0.h5` should not be used. It reuses the data from 6300_0 and does not hold measurements of case 7575. The tarball is left as published because repacking would make everyone download 16 GiB again, so exclude the file after you unpack. Only this one file is affected; the other angles of attack at Re 7575 are fine.

Physical setting

Flow around a NACA4418 airfoil cross-section. Real measurements are time-resolved Particle Image Velocimetry in a circulating tunnel; the simulated split is matched 3D CFD under the same geometry and operating conditions. Channels are [u, v, p], and p is zero on real data because it is not measured. Competition evaluation runs at 32 x 64 after 2x spatial subsampling; the raw PIV fields are 64 x 128.

example_data/3750_0.h5 stores u and v at the top level with shape (T=868, 64, 128), plus scalars aoa and re and the grids x, y. Note the training tarballs use a different layout, with u and v under measured_data/.

Baseline checkpoints

All were trained with the RealPDEBench code. Sizes: CNO 32 MB, Transolver 50 MB, FNO 403 MB in fp32 and 201 MB packed to fp16. Hyper-parameters should be read from the state_dict shapes rather than from the repository's yaml configs, which are not reliable for these checkpoints; the competition starting kit documents them.

FNO spectral weights are complex64, so a naive .half() breaks them. pack_ckpt_fp16.py stores them as view_as_real(t).half() with a complex_keys list, and the starting kit's load_baseline.unpack_fp16 reads that format back.

Using this in the competition

Competition rules require that training data come from this release. Model weights trained from it are permitted, including the baseline checkpoints above. Standard augmentation of the release is allowed, as long as every training sample traces back to a released sample through a transformation you can name. Generating additional data, using external data, and finetuning a model pretrained on anything else are not permitted, and the rule is checked when the shortlisted teams are re-trained from scratch before the final ranking. The Rules page and the FAQ of each track are authoritative.

License

Released for non-commercial research and competition use under CC BY-NC 4.0. Do not redistribute the hidden validation or test data, which is not in this release in any case.

Contact

realpde-competition@googlegroups.com, or the forum on either Codabench page.