Imagine a philharmonic orchestra playing, but the microphone picks up only a blended roar — a tutti where double basses, flutes, and timpani all merge. Physicists have grappled with a similar task for decades, staring at fields of density, temperature, or magnetic induction. In these multi-scale environments — from turbulent plasma to interstellar gas — nested structures lie hidden, but without a score, it’s unclear which instrument does what. Classical methods either average data into spectra, losing spatial organization, or impose templates like tracing filaments by gradients. What if a neural network learned the score by listening to the noise?

ScaleAware-JEPA acts like a virtuoso conductor who can “hear out” individual parts. First, the field undergoes diffusion decomposition (CDD) — the data is blurred over time, and each moment sets its own scale, much like an artist defocuses the background to bring large forms forward.

Then a neural network with a JEPA architecture learns to predict masked patches, not in raw pixels, but in a compact latent space — an abstract representation where similarity is defined by morphology, not geometry. The key twist: the size of the mask hiding the data is tied to the component’s scale. Small details need a tiny context (like a piccolo passage), while large ones demand a broad sweep (like a tuba’s slow theme). In this way, the network, akin to Riemann envisioning higher-dimensional spaces, weaves the local and global into a unified picture.

They tested on three examples. Numerical simulation of compressible magnetohydrodynamic turbulence, where plasma obeys Maxwell’s equations (unifying electricity and magnetism — a contribution of Maxwell) and Navier-Stokes equations. The network saw only density, yet its latent space, without hints, split into magnetic and kinetic structures — previously distinguishable only from magnetic field data. The second example: a map of night-time lights in Chengdu. The algorithm, unaware of the concept of “city,” on its own separated bright centers, roads, and diffuse outskirts. The third: molecular gas in the galaxy NGC 3627, observed with the ALMA telescope. Here a true latent atlas emerged: compact clouds — cradles of star formation, filamentary dust nebulae, and extended diffuse fields.

The main lesson: meaningful representations are born from data, not instructions. Where there are no ready-made phase diagrams, latent atlases become a compass. It’s striking that magnetic reconnection and turbulent cascades leave an imprint in ordinary density — many physical processes are secretly written into what we see. Who knows, might the algorithm find patterns not yet expressed in equations? Looking ahead: generalization to 3D data and time — static maps will turn into evolution movies. Merging with physics-informed networks (where Maxwell’s equations or conservation laws are embedded) will boost interpretability. The approach will take root not only in astrophysics but in climatology, material science, urban studies — wherever data appears as continuous fields.