Spinning Out of the Blue: AI Finds Chiral Superconductivity
08 May 2026, Yanjiang
A self-attention neural network discovers chiral p+ip superconductivity by minimizing energy from raw electron coordinates, revealing emergent topological order.
What if a machine, knowing nothing of Cooper pairs or Bogoliubov excitations, could deduce the existence of a chiral superconductor simply by staring at electron coordinates and turning an energy knob? That is not a thought experiment. A team led by Liang Fu at MIT, with Chun‑Tse Li at Academia Sinica, has shown that a self‑attention neural network can do exactly that, as reported in a preprint (arXiv:2509.03683). Given only an attractive interaction between electrons and the mandate to minimize energy, the AI found the chiral p‑wave superconducting state – the kind that breaks time‑reversal symmetry and hosts exotic edge modes – all on its own.
Staying on the subject of artificial intelligence in physics, the tool in question is a general‑purpose Fermi neural network equipped with permutation‑equivariant self‑attention. The network learns a many‑body wavefunction, mapping electron coordinates to a complex number whose squared magnitude gives the probability of that configuration. There is no explicit pairing term, no Bogoliubov transformation, no guess about the order parameter. It is a blank slate.
Think of it like an archaeologist who digs up a broken clay tablet and, simply by computing the shards’ most natural fit, stumbles upon a spiral inscription that no human had predicted. The AI’s attention mechanism draws in correlations between electrons – weighing how each particle influences every other – much like freezing water seeps into cracks and expands them, eventually shaping the material in ways nobody anticipated. Unlike the archaeologist, however, the network builds the quantum state from first principles; it does not interpret artifacts. It searches the space of possible wavefunctions and settles into the one that minimizes the energy.
The first hint that something special was happening came from the pair‑binding energy. When the network was trained on a modest number of fermions in a periodic box, a clear even‑odd oscillation emerged: systems with an odd count showed a negative binding energy, while even‑count systems showed a positive one – the very fingerprint of Cooper pairing, where one electron remains unpaired at the Fermi level while the rest pair off. The AI, unaware of BCS theory, had rediscovered the pairing instability.
But the real surprise lay deeper. The team projected the optimized wavefunction onto definite angular momentum sectors under four‑fold rotational symmetry. They found that the ground state was not symmetric at all; it lived primarily in sectors with angular momentum m = 1 and m = 3 – states that break time‑reversal symmetry and carry a chiral phase winding. In plain language, the AI had chosen a superconductor whose Cooper pairs orbit each other like miniature whirlpools, with the phase of the quantum wavefunction twisting by a full 2pi as you circle the origin in momentum space.
Speaking of symmetry breaking, the network’s discovery of a chiral state is a textbook example of emergent order. To verify it, the team computed the two‑body reduced density matrix – a diagnostic that reveals long‑range order. The leading eigenvector displayed that same 2pi winding in momentum space, consistent with chiral p+ip pairing. More tellingly, the pair correlation function exhibited off‑diagonal long‑range order: the probability of finding two electrons bound together persisted across the entire system, the hallmark of a superfluid of Cooper pairs.
It is akin to synchronizing an entire planetary body to one ‘time zone’: by projecting the wavefunction onto specific angular channels, the network locked into a coherent rotational state, revealing that the underlying order is not merely a correlation but a rigid phase structure that extends across the lattice. The network, of course, did not decide in any purposeful sense; it simply followed the gradient of the energy landscape until it reached the deepest valley.
The training itself was remarkably robust. Across different hyperparameters – batch sizes, learning rates, hidden dimensions, numbers of determinants – the optimization consistently converged to the same chiral state. The computational cost scaled roughly as the square of the particle number, a consequence of the quadratic complexity of self‑attention. Even for thirty‑one electrons, a system that pushes the limits of conventional exact diagonalisation, the GPU hours remained manageable, on the order of a few tens of thousands of iterations.
Beyond the technical achievement lies a deeper question. The network was not trained on any human knowledge of pairing mechanisms; it discovered a topological superconducting state purely through energy minimization. This raises an unsettling possibility: perhaps the standard repertoire of condensed matter theory – the symmetry‑based classification of order parameters – is incomplete, or at least biased by our own constructions. If a neural network, given only the raw Hamiltonian, can unearth a state as exotic as chiral p+ip, what else might it find in the vast wilderness of quantum many‑body systems?
I wonder if, a decade from now, we will look back at this moment as the turning point when AI became not just a tool but a genuine discoverer of physical law – a collaborator that sees patterns invisible to our intuition. It is a prediction I would be happy to see come true.
The phase of the wavefunction, after all, is a profoundly quantum construct: it has no classical analogue, yet it governs interference, topology, and the very existence of superconductivity. That an AI, guided only by the imperative of lowering energy, could grasp that phase and weave it into a coherent order is a testament to the irreducible mathematical beauty of the quantum world. The machine did not reason about Cooper pairs or gauge symmetry; it simply found that the universe loves a twist. And in that twist lies the future of discovery.
Yanjiang is an online editor of Loom Science
References
- Chun-Tse Li et al., Attention is all you need to solve chiral superconductivity, arXiv:2509.03683