disco-ad

Trajectories used in the ICML 2026 paper Test-time Generalization for Physics through Neural Operator Splitting (Serrano, Han, Oyallon, Ho, Morel).

Format

Each HDF5 file contains:

• trajectories: shape (N, T, C, *spatial), float32
• env_id: shape (N,), int64 — environment index for each trajectory
• env_params/*: optional metadata mapping env_id back to PDE coefficients

Load via the project's train_generic.py / test_generic.py scripts:

from train.train_generic import GenericHDF5Dataset
ds = GenericHDF5Dataset(["train.h5"])

Or using huggingface_hub:

from huggingface_hub import hf_hub_download
local = hf_hub_download(repo_id="sogeeking/disco-ad", filename="train.h5", repo_type="dataset")

Code: https://github.com/LouisSerrano/disco-ttg