Fig. 7: Out of distribution generalization on the generalized J1-J2 Heisenberg model.
From: Foundation neural-networks quantum states as a unified Ansatz for multiple hamiltonians

a Generalized J1-J2 Heisenberg model on a square lattice, with two couplings, J2L/J1 and J2R/J1. When J2L = J2R, the model reduces to the standard J1-J2 Heisenberg model. The FNQS is trained on configurations with frustration on only one diagonal and tested on configurations where both diagonals are frustrated. b Spin-spin correlation function on a 6 × 6 lattice of a FNQS trained on \({{{\mathcal{R}}}}=1000\) different realizations of frustration affecting only one of the two diagonals of a square lattice (see Fig. 7 for a schematic representation). The model is tested on two in-distribution points (J2L/J1, J2R/J1) = (0.0, 0.3) and (0.3, 0.0), and on the out-of-distribution point (J2L/J1, J2R/J1) = (0.3, 0.3), where both diagonals are frustrated. For reference, the exact results of the J1-J2 Heisenberg model at J2/J1 = 0.3 are also shown. Inset: The red line shows how the spin-spin correlations are ordered in the panel (b). c Out of distribution generalization of the square magnetization associated with Néel antiferromagnetic order for the FNQS (green circles) as a function of the frustration ratio J2/J1 on a 6 × 6 cluster. Exact results (blue line) are shown for reference.