The Well: 15TB of Physics Simulations

TL;DR

The Well is a 15TB collection of 16 numerical-simulation datasets spanning biology, fluid dynamics, acoustic scattering, magneto-hydrodynamics, extra-galactic fluids, and supernova explosions, packaged as a unified PyTorch interface (WellDataset) for training and benchmarking ML surrogate models on spatiotemporal physical systems. Datasets range from 6.9GB to 5.1TB each and can be downloaded locally or streamed from Hugging Face; the project ships a Hydra-based benchmark harness with reference implementations (e.g. FNO) on each dataset. Led by Polymathic AI in collaboration with Flatiron Institute, NYU, Cambridge, Princeton, CEA DAM, and others; published at NeurIPS 2024.

Key claims

The collection contains 16 datasets totalling 15TB drawn from domain scientists and numerical-software developers, covering biological systems, fluid dynamics, acoustic scattering, MHD of extra-galactic fluids, and supernova explosions [project page §1].
A unified WellDataset PyTorch interface lets users index any dataset by name and split (well_dataset_name, well_split_name) through the same DataLoader pipeline [project page §“Tap into the Well”].
Most datasets are mirrored on Hugging Face under polymathic-ai/ and can be streamed by setting well_base_path="hf://datasets/polymathic-ai/", though local download is advised for large training runs [project page §“Streaming from Hugging Face”].
The repository ships a Hydra-configured training script with reference surrogate models (e.g. FNO) and per-dataset configs, runnable as python train.py experiment=fno data=active_matter [project page §“Benchmark”].
The underlying NeurIPS paper is Ohana et al. 2024, “The Well: a large-scale collection of diverse physics simulations for machine learning” (NeurIPS, vol. 37, pp. 44989–45037) [project page §“Citation”; arXiv:2412.00568].

Method

The Well is a data + benchmarking artifact, not a model. Each of the 16 sub-datasets is a collection of numerical-simulation rollouts produced by domain-specific solvers (e.g. lattice Boltzmann, FV/FD PDE solvers, MHD codes, radiative-transfer codes) and stored in a common on-disk layout. The Python package wraps these in WellDataset, a torch.utils.data.Dataset subclass keyed by (base_path, dataset_name, split_name) so that a single training loop can iterate any dataset without dataset-specific glue. The benchmark side wires WellDataset into a Hydra config tree alongside model definitions (the_well/benchmark/models, with FNO and other surrogates pre-implemented) and a train.py entry-point; switching dataset, model, or optimizer is a single CLI override. The HF-streamed path uses the same interface with an hf:// base URI.

Results

No headline numbers — this is a dataset release, not a model release. The benchmark scaffold is intended for downstream researchers to report FNO / U-Net / Transformer surrogate skill on each sub-dataset under matched training compute. The NeurIPS paper (arXiv:2412.00568) is the place to look for reference baselines on the 16 sub-datasets.

Why it’s interesting

The Well is the closest analog to ImageNet/LAION for physical-system surrogate learning: a heterogeneous, large-scale, unified-interface corpus aimed at training general-purpose neural surrogates for PDE/ODE systems. For a Luma context this is relevant alongside Skillful joint probabilistic weather forecasting from marginals — both are about training neural models on numerical-simulation data, and FGN’s marginals-only-CRPS-on-shared-conditional-norm recipe is exactly the kind of architectural trick a benchmark like The Well could surface or refute across many physical regimes. It also pairs naturally with world-model work like The Waymo World Model: A New Frontier For Autonomous Driving Simulation and Project Genie: Experimenting with infinite, interactive worlds: those models generate worlds without explicit physics; The Well is the dataset you’d use to ask whether their latents recover known PDE solutions, or to retrain a physically-grounded surrogate as a perception backbone.