shifushuimenu / autoregressive_Slater-Jastrow

Autoregressive Slater-Jastrow ansatz

Autoregressive Slater-Jastrow ansatz for variational Monte Carlo arXiv:2210.05871

---

As a demonstration the t-V model of spinless fermions on the square lattice is implemented.

Key features

• Get rid of autocorrelation time completely !
• Fast direct sampling from a Slater determinant
• Jastrow factor represented by autoregressive neural network
• Lowrank update for local kinetic energy preserves cubic scaling

How to run the code

Run VMC for t-V model on a square 4x4 lattice with 8 spinless fermions and interaction strength V/t=4.0; training iterations = 1000; batch size per iteration = 200; num. of samples in measurement phase = 300.

python3 -O ./run_autoregVMC.py 4 4 8 4.0 1000 200 300 --optimizer Adam --seed 42 --optimize_orbitals True

For a full list of options see

python3 ./run_autoregVMC.py --help

Benchmarking

python3 ./ED/ED_spinless_fermions_tV_2d.py 4 4 8 4.0

This requires installation of the QuSpin library.

Warning

If run in debug mode (i.e. without -O flag) the lowrank update on rare occasions throws an error due to limited floating point precision. This could be avoided using mpmath infinite precision library.

Cite

@article{humeniuk2022autoregressive,
  title={Autoregressive neural Slater-Jastrow ansatz for variational Monte Carlo simulation},
  author={Humeniuk, Stephan and Wan, Yuan and Wang, Lei},
  journal={arXiv preprint arXiv:2210.05871},
  year={2022}
}