Validation and cross-code parity

vibe-qc implements every method it ships (CLAUDE.md sec 10); established codes are used only as out-of-process reference oracles (parsed via the examples/regression/runner_<program>.py subprocess pattern, never imported into the runtime). This page consolidates the validation surface that the release paper draws on. Numbers are produced by the in-repo regression suite; per-system deltas are re-confirmed against the tagged release build.

Methodology

  • Same-family oracles. Each route is compared against the established code in its own method family, so a discrepancy isolates a real difference rather than a cross-family convention mismatch: periodic Gaussian-AO HF/DFT vs CRYSTAL (BIPOLE) and PySCF (GDF); plane-wave-grid DFT vs CP2K / GPAW (GPW/GAPW); molecular HF/DFT/MP2/CC vs PySCF; correlated/CC accuracy cross-checked vs ORCA; multireference vs OpenMolcas.

  • Matched settings. Cross-code comparisons fix the convergence-relevant knobs (k-mesh / SHRINK, integral tolerances, smearing, mixing) so the residual reflects the method, not the protocol.

Molecular methods (vs PySCF / ORCA)

  • HF / DFT / MP2 / gradients / Hessian – machine precision vs PySCF (see docs/features.md).

  • Canonical CCSD(T) (RHF / UHF / ROHF references) – validated to uHa vs PySCF cc.CCSD/CCSD(T) and cc.UCCSD/UCCSD(T), with ORCA 6.1 cross-checks.

  • DLPNO-CCSD(T) – defaults give MAE 0.30 kcal/mol vs canonical CCSD(T) (0.166 with tcut_pno=1e-8 + triples_mode="exact"); full-domain reduces exactly to canonical; cross-validated vs ORCA 6.1. Always report DLPNO accuracy with its threshold – the number is meaningless without tcut_pno.

  • DLPNO-MP2 (closed- and open-shell UMP2) – defaults recover >= 99.8 % of canonical RI-MP2 correlation energy; thresholds -> 0 reproduce it to <= 1 uHa.

  • Multireference (CASCI / CASSCF / NEVPT2 / IC-CASPT2) – validated vs PySCF (mcscf, mrpt) and OpenMolcas where PySCF has no reference.

Periodic HF / DFT (vs CRYSTAL / PySCF / CP2K)

Periodic Hartree-Fock and Kohn-Sham DFT are vibe-qc’s distinctive contribution: three independent in-house routes, each validated against the established code in its family. (Periodic post-HF correlation is out of scope here – a separate study.)

  • BIPOLE (CRYSTAL-family). Bipolar/multipolar real-space + Ewald exchange. Validated against CRYSTAL at matched k-mesh and tolerances via the examples/regression/crystal_parity/ harness (local CRYSTAL binary, out-of-process).

  • GDF (PySCF-family). Gaussian density fitting, the default run_periodic_job backend; uHa parity vs PySCF KRHF.density_fit / KRKS.density_fit at matched k-mesh.

  • GPW / GAPW (CP2K / GPAW-family). Gaussian-and-plane-waves; validated against CP2K / GPAW reference runs.

CRYSTAL14 reference targets (STO-3G, SHRINK 8 8, TOLDEE 8; per formula unit)

These are the sealed CRYSTAL14 references the BIPOLE/GDF periodic-HF parity tests match (examples/regression/crystal_parity/baseline_sto3g/PARITY_TABLE.md):

System

E_total/FU (Ha)

LiH

-7.93816810

MgO

-271.21814708

NaCl

-614.65306571

LiF

-105.63898757

C-diamond

-74.87695367

Si-diamond

-571.31994169

Where the numbers live

The validation is executed, not asserted by hand: the molecular cross-checks are pinned in the tests/ suite (PySCF/ORCA references with inline provenance), and the periodic parity is driven by examples/regression/crystal_parity/ and examples/regression/bipole_parity/. The full-suite gate (scripts/test_gate/) keeps these green on every release.