The Well - A 15TB Physics Simulation Dataset Collection

"The Well" is a large-scale collection of physics simulation datasets, totaling 15TB, designed for use in machine learning research, particularly for training and evaluating surrogate models. The project aims to address the limitations of existing datasets that often cover only narrow classes of physical behavior. It provides a diverse range of spatiotemporal simulations from various domains, making it a valuable benchmark for testing the generalization capabilities of machine learning models in physics-related tasks.

Addressing the Data Gap:

The primary motivation behind "The Well" is to overcome the lack of comprehensive and diverse datasets in the field of physics-informed machine learning. The project creators state, "as standard datasets in this space often cover small classes of physical behavior, it can be difficult to evaluate the efficacy of new approaches."

Diversity of Simulations:

The dataset collection spans a wide array of physical phenomena, including "biological systems, fluid dynamics, acoustic scattering, as well as magneto-hydrodynamic simulations of extra-galactic fluids or supernova explosions." This breadth ensures that models trained on "The Well" are challenged with complex and varying dynamics. The 16 datasets offer a variety of complexity.

Benchmarking & Evaluation:

"The Well" is explicitly designed as a benchmark suite. The repository provides not just data but also tooling and baseline models to facilitate the evaluation of new algorithms and architectures, enabling a comparison of methods on a common ground.

Open-Source and Accessible:

The data and code are openly available on GitHub, making "The Well" accessible to a broad community. The team emphasizes a unified PyTorch interface to streamline training and model evaluation. "To facilitate usage of the Well, we provide a unified PyTorch interface for training and evaluating models."

Collaboration:

The project is a collaborative effort involving researchers from multiple institutions, highlighting its interdisciplinary nature and the importance of shared resources for advancing machine learning in scientific computing.

"This project has been led by the Polymathic AI organization, in collaboration with researchers from the Flatiron Institute, University of Colorado Boulder, University of Cambridge, New York University, Rutgers University, Cornell University, University of Tokyo, Los Alamos Natioinal Laboratory, University of California, Berkeley, Princeton University, CEA DAM, and University of Liège."

Numerical Simulations:

The data originates from numerical simulations, rather than real-world measurements. This focus enables researchers to test a variety of physical phenomena without the limitations of physical experimentation.

Surrogate Modeling:

The primary use case for "The Well" is in developing surrogate models that can emulate computationally expensive simulations.

Key Facts:

• Size: The total size of "The Well" is 15TB. The individual datasets range in size from 6.9GB to 5.1TB.
• Number of Datasets: There are 16 individual datasets included in the collection.
• Domains Covered: The datasets span biological systems, fluid dynamics, acoustic scattering, and magneto-hydrodynamics.
• Programming Interface: "The Well" provides a PyTorch interface for data access and model training.
• Availability: The data is available through the GitHub repository and also streamed through Hugging Face. The github repository can be found here: https://github.com/PolymathicAI/the_well.
• Software Installation: The "the_well" python package can be installed via pip, either from the PyPI repository or directly from source.
• Citation: The project has been documented in a research paper titled, "The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning," presented at the NeurIPS 2024 Datasets and Benchmarks Track. The Bibtex entry is provided for citation.

Usage and Implementation:

• Installation: Users can install the "the_well" Python package using pip install the_well.
• Data Download: The datasets can be downloaded using the the-well-download command-line tool.
• Streaming: The data can be streamed directly from Hugging Face with the 'hf://datasets/polymathic-ai/' URL using the WellDataset class.
• Benchmarking: The repository includes scripts and configuration files for running benchmark models on the datasets, using python train.py.
• Customization: Users can override or edit configuration using YAML files, allowing for flexibility in model and training parameters.

Conclusion:

"The Well" represents a significant contribution to the field of machine learning in physical sciences. By providing a large, diverse, and accessible collection of physics simulation datasets, this project has the potential to accelerate research on surrogate models, operator learning, and other related areas. The well-structured code, benchmark tools, and clear documentation contribute to its utility and make it a promising resource for both researchers and practitioners. The project encourages feedback and contributions via the GitHub repository, fostering a collaborative environment.