PolymathicAI Releases 'The Well': A 15TB Benchmark Suite of Physics Simulation Datasets
Key Takeaways
- ▸15TB of physics simulation data across 16 datasets covering diverse domains from biology to astrophysics, usable individually or as a complete benchmark suite
- ▸Multi-channel distribution (GitHub, PyPI, Hugging Face) with options to download locally or stream directly, plus data ranging from 6.9GB to 5.1TB per dataset
- ▸Includes benchmarking infrastructure, pre-trained models, and training scripts for evaluating surrogate model performance across different physics simulations
Summary
PolymathicAI has released The Well, an open-source collection of 15TB of machine learning datasets containing numerical simulations of diverse physical systems. The dataset suite comprises 16 datasets spanning biological systems, fluid dynamics, acoustic scattering, magneto-hydrodynamic simulations, and supernova explosions—designed to accelerate research in machine learning and computational sciences.
Individual datasets range from 6.9GB to 5.1TB and are available through multiple distribution channels: GitHub, PyPI, and Hugging Face. The package includes a Python API for seamless integration into training pipelines, along with benchmarking infrastructure and pre-implemented state-of-the-art surrogate models like Fourier Neural Operators (FNO) for model evaluation.
The release bridges machine learning and computational science by providing standardized physics simulation data that researchers can use to train and evaluate deep learning models. Documentation includes tutorials, API references, and training scripts via Hydra configuration, enabling users to benchmark various architectures across multiple physics domains.
Editorial Opinion
The Well fills a critical gap in physics-informed machine learning by providing researchers with standardized, large-scale simulation datasets across multiple domains. By open-sourcing both data and training infrastructure, PolymathicAI has created a valuable platform for developing generalizable surrogate models that could accelerate scientific discovery. This kind of domain-specific benchmark suite is essential for advancing physics-informed AI and bridging the gap between domain science and machine learning communities.
