Flow-based sampling for fermionic lattice field theories
Paper • 2106.05934 • Published
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Scalar-field sector of the 2D staggered Yukawa model from Albergo et al., arXiv:2106.05934, sampled with pseudofermion HMC, plus a score-based diffusion model (VE-SDE, NCSN++-style 2D U-Net) trained on the g=0.1 ensemble.
16x16 lattice, two-flavor staggered fermions, m_f = 0.
| set | m^2 | lambda | g | <|M|> (HMC) |
|---|---|---|---|---|
| g=0.1 | -4.00 | 6.0 | 0.1 | 0.07326(17) |
| g=0.3 | -1.55 | 2.4 | 0.3 | — |
YukawaFermionHMC2D.jl, generate_samples.jl, force_ratio.jl — Julia
pseudofermion HMC (sparse Cholesky solves; paper Table II force-ratio checks)samples/yukawa_g{0.1,0.3}_L16_1000000.jld2 — 100k scalar configs each
(key configs, shape (100000, 16, 16); 1M trajectories, save_every=10),
plus .npz mirrors and per-trajectory |M| historiesdiffusion/ — PyTorch training/sampling/analysis for the diffusion model
(train_yukawa.py, sample_yukawa.py, analysis_observables.py,
plot_chi_hist.py; network/SDE definitions live in the parent DM repo)diffusion/runs/yukawa_L16_g0.1_ncsnpp/models/ — log-spaced checkpoints
(epochs 1–65, sigma=50, batch 256, lr 1e-3, bf16)diffusion/runs/yukawa_L16_g0.1_ncsnpp/data/ — generated samples
((16, 16, N) npy, physical field units) and comparison figures| observable | HMC (N=100k) | diffusion model |
|---|---|---|
| <|M|> | 0.07326(17) | 0.07328(59) |
| chi = V(<M^2> - <|M|>^2) | 0.7469(36) | 0.759(14) |