Polymathic AI Released 15TB of Real Physics Simulations and It Changes How We Train Scientific Models

Polymathic AI has released "The Well," an open-source library containing 15 terabytes of real physics simulations spread across 16 datasets. The release represents a fundamental shift in how AI models learn about the physical world. Instead of training on text descriptions of physical phenomena, models can now learn directly from simulation data covering supernovas, fluid dynamics, biological systems, and more.

What The Well Actually Contains

The Well is a large-scale collection of machine learning datasets built from numerical simulations of spatiotemporal physical systems. Domain scientists and numerical software developers contributed 15TB of data across 16 distinct datasets. Those datasets cover fluid dynamics, acoustic scattering, biological systems, and hydrodynamic simulations of extra-galactic fluids and supernova explosions.

The project is hosted on GitHub at PolymathicAI/the_well, published on PyPI as version 1.2.0, and available through Hugging Face as downloadable datasets. The accompanying research paper (arXiv: 2412.00568) was presented at NeurIPS 2024, one of the top machine learning conferences.

Why This Matters: From Generative AI to Scientific AI

The core argument in this video is about a category shift. Most AI models learn about physics the way a person might learn to swim by reading a book about water. They process text descriptions and statistical patterns, but they never interact with the actual physical data.

The Well changes that equation. Models can now experience physical phenomena directly through real simulation data. This is a move from generative AI, which produces plausible-sounding outputs based on pattern matching, toward scientific AI, which can discover and validate real physical truths. That distinction matters because scientific applications demand accuracy over fluency.

Who Benefits From This

Researchers working on physics-informed machine learning get the most immediate value. Training a neural network to predict fluid flow, for example, requires massive amounts of simulation data that individual labs often cannot generate on their own. The Well consolidates that data into one accessible, well-documented source.

The BSD-3-Clause license means the datasets are free to use in both academic and commercial contexts. Open access to this scale of scientific data lowers the barrier for smaller research groups and independent developers who want to build physics-aware AI systems without running their own supercomputer simulations.

Key Takeaways

• The Well provides 15TB of physics simulations across 16 datasets, covering domains from fluid dynamics to supernova explosions
• Models train on real simulation data instead of text-based descriptions of physical phenomena
• The project is fully open source under a BSD-3-Clause license, with code on GitHub and data on Hugging Face
• This was peer-reviewed work presented at NeurIPS 2024, giving it stronger credibility than a typical open-source release
• The practical shift is from generative AI to scientific AI, where models discover physical truths rather than generating plausible text

Resources

• The Well GitHub Repository -- source code, documentation, and dataset links
• arXiv Paper (2412.00568) -- the full research paper behind the project
• Hugging Face Datasets -- download the simulation datasets directly