PolymathicAI/the_well

What it solves

It provides a massive, standardized collection of numerical physics simulations to accelerate the development of machine learning models for computational sciences. By offering 15TB of diverse spatiotemporal data, it removes the hurdle of sourcing high-quality simulation data for training surrogate models that can predict physical systems.

How it works

The project consists of a Python package that allows users to download, stream (via Hugging Face), and load physics datasets into training pipelines using a WellDataset class compatible with PyTorch. It includes 16 datasets covering fields like fluid dynamics, biological systems, and supernova explosions. Additionally, it provides a benchmarking suite with pre-implemented surrogate models (such as FNO) and training scripts to evaluate model performance on these physical systems.

Who it’s for

Researchers and machine learning engineers working on PDE (Partial Differential Equation) surrogate modeling and computational physics who need large-scale, diverse training data.

Highlights

• Massive Scale: Contains 15TB of data across 16 diverse physics datasets.
• Broad Domain Coverage: Includes simulations of everything from acoustic scattering to magneto-hydrodynamic extra-galactic fluids.
• Flexible Data Access: Supports both local downloads and direct streaming from Hugging Face.
• Integrated Benchmarking: Includes a training framework and pre-trained baseline model checkpoints for easy comparison.