In the quantum world, particles fall into two irreconcilable camps: bosons love to pile into the same state (like light in a laser), while fermions are individualists that never occupy the same spot. You'd think one neural network couldn't learn both personalities at once. But the authors found a loophole: they taught the network to pretend. Meet ChernFormer — an architecture that describes bosons but thinks like fermions.

The trick is to attach a fixed "twist" to each pair of particles — the so-called Chern–Simons phase (a purely mathematical add-on that doesn't change the probability of finding particles but flips the sign of the wave function upon exchange). The fermionic part of the network changes sign when particles are swapped, and the added twist also changes sign. Two minuses make a plus, so the whole construction behaves like a boson. This way, any bosonic ensemble can be recast as a fermionic one with the same accuracy.

The authors tested the network on a challenging problem — reconstructing the ground state and two edge excitations of the Kalmeyer–Laughlin topological liquid (a state of matter where particles collectively behave as a single entity with fractional charges). For 20 particles, ChernFormer accurately reproduced both the local vortices (density swirls around each particle) and the collective edge vortex — something ordinary bosonic networks miss. The overlap with the reference remained close to unity for all tested sizes. Moreover, if the network is allowed to smooth out the "hole" at zero (the contact singularity when particles approach), it smoothly transitions into a description of an ordinary condensate with no nodes at all.

So one tool covers two extremes of the quantum world. Previously, each type of order required its own approach; now there's a common variational language. This is a step toward universal simulators of quantum matter, where the neural network itself figures out what it's dealing with: a peaceful condensate or the intricate weave of topological edges.