Restricted Boltzmann machine network versus Jastrow correlated wave function for the two-dimensional Hubbard model

TL;DR

This paper demonstrates that a neural network-based RBM wave function provides a superior variational approach for the two-dimensional Hubbard model, especially in capturing antiferromagnetic correlations, compared to traditional Jastrow methods.

Contribution

It introduces a neural network-based RBM wave function for the 2D Hubbard model and shows its advantages over Jastrow wave functions in describing magnetic properties.

Findings

RBM wave function outperforms Jastrow in variational energy in the underdoped region.

RBM spontaneously exhibits strong antiferromagnetic correlations.

RBM provides an improved phase diagram description of the model.

Abstract

We consider a restricted Boltzmann Machine (RBM) correlated BCS wave function as the ground state of the two-dimensional Hubbard model and study its electronic and magnetic properties as a function of hole doping. We compare the results with those obtained by using conventional Jastrow projectors. The results show that the RBM wave function outperforms the Jastrow projected ones in the underdoped region inmterms of the variational energy. Computation of superconducting (SC) correlations in the model shows that the RBM wave function gives slightly weaker SC correlations as compared to the Jastrow projected wave functions. A significant advantage of the RBM wave function is that it spontaneously gives rise to strong antiferromagnetic (AF) correlations in the underdoped region even though the wave function does not incorporate any explicit AF order. In comparison, AF correlations in the Jastrow projected wave functions are found to be very weak. These and other results obtained show that the RBM wave function provides an improved description of the phase diagram of the model. The work also demonstrates the power of neural-network quantum state (NQS) wave functions in the study of strongly correlated electron systems.