engineering: ph-AFQMC post-processing using SQD as trial wavefunction (literature-validated path to chem-acc on N2-CAS(10,26))

[thc1006]

TL;DR

Mahajan et al. JCTC 2025 demonstrated that phaseless AFQMC using truncated SQD wavefunction as trial recovers O(100) mHa correlation energy for N2-CAS(10,26) — exactly our 52Q target system. Even random LUCJ samples work. Does not require wavefunction tomography. This is the literature-validated chem-acc path on our specific system.

Why this matters now

Our HI-NQS-SQD pipeline plateau at 52Q ≈ -109.245 Ha. Literature shows ph-AFQMC + SQD trial can reach chem-acc (within 20 mHa from FCI reference). We're hitting a method ceiling, not a hyperparameter ceiling.

Reference (5+ verified sources for AFQMC + SQD line)

• Mahajan et al. (JCTC 2025, 5c01407) — direct N2-CAS(10,26) benchmark
• arXiv:2503.05967 — preprint of above
• NNQS-AFQMC (arXiv:2507.07540) — NQS+AFQMC parallel work
• Beyond CCSD(T) Accuracy with AFQMC (JCTC 2024, 4c01314)
• On AFQMC's transformative role (PMC 2023)

Proposal

1. Add qvartools.solvers.afqmc.PhaselessAFQMCSolver interface
2. Use existing ipie library (BSD-licensed, pyscf-compatible) as backend
3. Take SolverResult from HI-NQS-SQD as input (truncated SCI wavefunction → AFQMC trial)
4. Run ph-AFQMC propagation (typically 200 walkers × 1000 steps ≈ 30 min)

sqd_result = run_hi_nqs_sqd(...)
afqmc_result = run_phaseless_afqmc(
    trial=sqd_result.sci_state,
    n_walkers=2000,
    n_steps=200,
    constraint="phaseless",
)
# Expected: -50 to -100 mHa beyond SQD on 52Q

What this is NOT

This is engineering, not research. ph-AFQMC + SQD trial is published methodology. We adopt it for validation/production, not novelty.

Effort: 2-3 weeks

• ipie integration
• API design (SolverResult → AFQMC trial conversion)
• Tests on small molecules (BeH2, H2O) before 52Q
• Documentation