Driven solely by the density of interstellar gas, a network rediscovers the magnetohydrodynamic organization that had been hidden from it

Velocity and magnetic field are removed from the drive, then reintroduced as diagnostics. The geometry learned on morphology is however ordered according to these quantities.

The morphology of turbulent interstellar gas results from the coupled action of turbulence, magnetic fields and self-gravity. But density amplitude, taken alone, is not enough to determine the dynamic state of a region — two structures of comparable density may have very different regimes.

The authors induce ScaleAware-JEPA to learn multi-scale structural coordinates from density only and then check whether these coordinates order dynamic states better than direct conditioning on density. Latent neighbourhoods corresponding to logs, filaments, and diffuse gas show an orderly progression of turbulent velocity scattering, which populations selected on density do not reproduce.

The protocol is the interesting part: speed and magnetic field are deliberately excluded from training, and then used after it as physical diagnostics. The magnetohydrodynamic states then occupy orderly, though partially superimposed, regions of the observed representation, and dynamic equilibriums without dimension vary coherently over the whole latent atlas. In other words, a representation built on the density geometry alone acquires a structure that established MHD diagnostics systematically organize.