The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning
Summary
The Well is a large-scale collection of 15TB of diverse physics simulation datasets across 16 domains, designed to benchmark machine learning surrogate models for spatiotemporal physical systems. It provides a unified PyTorch interface and example baselines to accelerate simulation-based workflows.
View Cached Full Text
Cached at: 06/27/26, 05:18 PM
Paper page - The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning
Source: https://huggingface.co/papers/2412.00568 Published on Nov 30, 2024
Authors:
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
Abstract
A large-scale dataset collection, The Well, provides diverse numerical simulations for benchmarking machine learning models in physical systems simulation.
Machine learning basedsurrogate modelsoffer researchers powerful tools for acceleratingsimulation-based workflows. However, as standard datasets in this space often cover small classes of physical behavior, it can be difficult to evaluate the efficacy of new approaches. To address this gap, we introduce the Well: a large-scale collection of datasets containingnumerical simulationsof a wide variety ofspatiotemporal physical systems. The Well draws from domain experts and numerical software developers to provide 15TB of data across 16 datasets covering diverse domains such asbiological systems,fluid dynamics,acoustic scattering, as well asmagneto-hydrodynamic simulationsof extra-galactic fluids or supernova explosions. These datasets can be used individually or as part of a broaderbenchmark suite. To facilitate usage of the Well, we provide a unifiedPyTorch interfacefor training and evaluating models. We demonstrate the function of this library by introducing example baselines that highlight the new challenges posed by the complex dynamics of the Well. The code and data is available at https://github.com/PolymathicAI/the_well.
View arXiv pageView PDFGitHub3.63kAdd to collection
Get this paper in your agent:
hf papers read 2412\.00568
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2412.00568 in a model README.md to link it from this page.
Datasets citing this paper13
#### polymathic-ai/acoustic_scattering_inclusions UpdatedApr 10, 2025 • 23.5k #### polymathic-ai/rayleigh_benard UpdatedApr 10, 2025 • 6.88k #### polymathic-ai/planetswe UpdatedApr 10, 2025 • 5.87k • 1 #### polymathic-ai/acoustic_scattering_discontinuous UpdatedApr 10, 2025 • 5.34k • 1 Browse 13 datasets citing this paper### Spaces citing this paper1
Collections including this paper5
Similar Articles
@HowToPrompt__: someone open-sourced 15TB of physics simulations that would take a national lab and millions in supercomputer time to r…
The Well is an open-source collection of 15TB of physics simulation datasets spanning 16 domains, created by Flatiron Institute and 11 universities, designed to train PDE surrogate models, enabling researchers to replace expensive supercomputer simulations with neural networks.
@Vincent_AINotes: Just now, the research community was blown away by an open-source project called The Well. Someone packaged high-precision physics simulation results that would normally require a national lab, millions of dollars, and weeks on a supercomputer, totaling 15TB, and threw it online, 100% free and open source! This is the "digital twin" collection of humanity's top computing power...
The Well is an open-source project providing 15TB of high-precision physics simulation results, which would normally cost millions of dollars and weeks on a supercomputer at a national lab, now 100% free and open source, making a huge impact on the scientific research field.
ThousandWorlds: A benchmark for climate emulation of potentially habitable exoplanets
ThousandWorlds is a benchmark dataset for machine-learning emulation of exoplanet climates, containing approximately 1800 simulations from five global climate models. Gaussian process methods outperform deep learning baselines in this low-data, multi-simulator regression task.
@heyrobinai: THE ENTIRE AI INDUSTRY JUST GOT HUMILIATED a tiny model trained in just a few hours on a single graphics card is planni…
Yann LeCun's team releases LeWorldModel, a tiny 15M-parameter physics model trained on a single GPU in hours that outperforms billion-dollar foundation models in planning speed and physical plausibility, challenging the dominant scaling paradigm.
The State of Simulation for Physical AI: An Overview
This blog post by NVIDIA provides an overview of simulation engines for physical AI, covering why simulation is critical for robotics data generation and comparing tools like MuJoCo, Isaac Sim, and Newton. It highlights the shift from simulation as a debugging tool to an integral part of model development.