Convergence to the fixed-node limit in deep variational Monte Carlo

Variational quantum Monte Carlo (QMC) is an ab-initio method for solving the electronic Schr\"odinger equation that exact in principle, but limited by flexibility of available ansatzes practice. The recently introduced deep QMC approach, specifically two deep-neural-network PauliNet and FermiNet, allows variational to reach accuracy diffusion QMC, little understood about convergence behavior such ansatzes. Here, we analyze how approaches fixed-node limit with increasing network size. First, demonstrate a neural can overcome limitations small basis set mean-field complete-basis-set limit. Moving electron correlation, then perform extensive hyperparameter scan Jastrow factor LiH H$_4$ find energies at be obtained sufficiently large network. Finally, benchmark many-body on H$_2$O, fraction recovered correlation energy single-determinant Slater--Jastrow-type half order magnitude compared previous results Slater--Jastrow--backflow version ansatz overcomes limitations. This analysis helps understanding superb comparison traditional trial wavefunctions respective level theory, will guide future improvements architectures QMC.