Skip to main content
Back to timeline
arXivSource publication:

KagomeResNet pushes the kagome spin-liquid calculation to 300 spins, and the ground state stays at Γ while three diagnostics point to a Dirac spin liquid

Synopsis

For the spin-1/2 kagome Heisenberg antiferromagnet, whose ground state has long been unresolved, KagomeResNet embeds message passing along nearest-neighbour bonds into a neural-network quantum state and reaches exceptionally low variational energies on two-dimensional tori of up to 300 spins. Spin and dimer structure factors show no sign of magnetic or valence-bond long-range order, the dominant symmetry-resolved correlations share a proportional scaling consistent with the monopole multiplet of a Dirac spin liquid, and the optimized wave function's fidelity to a Gutzwiller-projected Dirac state increasingly favours it over competing candidates as the system grows.

Source-provided article image: Evidence for a Dirac Spin Liquid in the Kagome Heisenberg Antiferromagnet from High-Precision Neural-Network Quantum States
arXiv

Interpretation

The ground state of the spin-1/2 kagome Heisenberg antiferromagnet is placed in the Γ-point, zero-total-spin sector, and no size-driven change of ground-state quantum numbers appears. An earlier machine-learning study based on a group convolutional network reported that the ground-state sector moved from Γ to K with size and proposed a spinon pair-density-wave ground state on that basis. On tori of 6, 8 and 10 unit cells (108, 192 and 300 spins), the lowest state in the odd spin-inversion sector sits at K, but comparing both sectors leaves the overall ground state at Γ, matching the quantum numbers exact diagonalization gives for 6 unit cells.

The optimized neural-network quantum state shows no long-range order in either the spin or the dimer channel, and its dominant fluctuations fall in specific symmetry-resolved channels. Earlier projected mean-field variational calculations could not separate the U(1) Dirac state from competing spin-liquid states because their variational energies are nearly degenerate. The spin structure factor has no sharp peak inside the first Brillouin zone and its normalized maximum tends to zero with size; the dimer structure factor likewise does not grow with size, and the dominant dimer fluctuation belongs to the mirror-odd channel.

The symmetry-resolved Γ-point spin correlations and K-point dimer correlations collapse onto a single proportional relation across distances and the three system sizes. These two correlations carry the quantum numbers of the spin-triplet and spin-singlet minimal monopoles in the QED3 description of the Dirac spin liquid, which theory expects to share one scaling dimension. Data from the three system sizes fall on the same proportional function, the behaviour expected for different components of the same emergent monopole multiplet.

As the system grows, the optimized neural-network quantum state's fidelity to a Gutzwiller-projected Dirac reference increasingly exceeds that to competing candidates. Within the projected mean-field framework the variational energies of the Dirac and competing states are nearly indistinguishable, so energy ordering alone cannot settle the ground state. Fidelity to the Dirac reference is significantly higher than to a finite-pairing reference whose energy is indistinguishable from it; against an optimized resonating-valence-bond reference with slightly lower energy, the two are close at 6 unit cells and the fidelity ratio grows linearly with size thereafter.

Perspective

The conclusion concerns the spin-1/2, nearest-neighbour, periodic-boundary kagome Heisenberg model, computed on two-dimensional tori with 6, 8 and 10 unit cells, that is 108, 192 and 300 spins. The authors propose extending KagomeResNet to the long-range kagome model and to ground states of more complex lattices such as the shuriken and maple-leaf lattices.

The Dirac spin-liquid identification rests on extrapolating finite-size data, with the correlation scaling fitted on the three sizes of 6, 8 and 10 unit cells. The authors also note that previously reported unusually low variational energies can arise from nonergodic Monte Carlo sampling, so ergodicity checks are a prerequisite for interpreting such neural-network quantum-state results.