Multiple Physics Pretraining

One transformer learns the dynamics of many different equations at once and adapts to systems it has never seen.

Made by
Michael McCabe, Bruno Régaldo-Saint Blancard, Liam Parker and others (Polymathic AI)
Field
Simulation and fluids
Learns from
Fluid simulations of several physical systems
Size
7.6M to 116M parameters
Open weights
Yes, MIT (code)
First public
Oct 2023

How it relates to Flare

Early evidence that pretraining across physics pays off, the idea CERN and Mistral say could later reach other sciences.

Every figure here is from the source above, checked 17 October 2026. Something changed? Email [email protected].

All physics AI models compared