"""Cell definitions for the ORCA parity matrix + speed benchmark. A *cell* is one ``(system, basis, method, charge, spin, df, cosx)`` tuple — the unit both deliverables iterate over. Geometries are written here in Angstrom (human-readable) and converted **once** to Bohr by :func:`geometry_bohr_symbols`; both the ORCA input (fed Bohr via the ``Bohrs`` keyword) and the vibe-qc side consume that single conversion, so the two codes sit on identical nuclear coordinates. The parity geometries mirror ``tests/test_parity_hf_dft.py`` exactly (H2O / H2CO / OH at the same coordinates) so the ORCA axis lines up cell-for-cell with the PySCF axis already in that file. The speed ladder adds larger systems to find the DF / RIJCOSX crossover the handover calls for. """ from __future__ import annotations from dataclasses import dataclass from typing import List, Tuple # --- geometries (Angstrom) ---------------------------------------------- # H2O / H2CO / OH are the exact coordinates of test_parity_hf_dft.py's # _GEOMS, just expressed in Angstrom instead of Bohr. GEOMETRIES = { "H2O": [ ("O", (0.0, 0.0, 0.0)), ("H", (0.0, 0.793353, -0.613510)), ("H", (0.0, -0.793353, -0.613510)), ], "H2CO": [ ("C", (0.0, 0.0, 0.0)), ("O", (0.0, 0.0, 1.205)), ("H", (0.0, 0.943, -0.587)), ("H", (0.0, -0.943, -0.587)), ], "OH": [ ("O", (0.0, 0.0, 0.0)), ("H", (0.0, 0.0, 0.97)), ], # Open-shell radicals — same geometries as # `tests/test_parity_hf_dft.py::_GEOMS` (NIST CCCBDB exp bond # lengths for the diatomics; published B3LYP/aug-cc-pVTZ for # methylperoxyl). BRIEF.md § Scope-2 coverage. "NO": [ ("N", (0.0, 0.0, 0.0)), ("O", (0.0, 0.0, 1.151)), # r(N-O) = 1.151 Å ], "CN": [ ("C", (0.0, 0.0, 0.0)), ("N", (0.0, 0.0, 1.172)), # r(C-N) = 1.172 Å ], # Methylperoxyl CH3O2· — Cs, doublet. Constructed from r(C-O)=1.45, # r(O-O)=1.32, r(C-H)=1.094 Å with ∠(C-O-O)=110°, ∠(O-C-H)=109.5° # tetrahedral, H atoms staggered (60°/180°/300° azimuthal) so the # anti-H sits anti-periplanar to the radical site. Coordinates # below are this construction expressed in Å. "CH3OO": [ ("C", ( 0.0000000, 0.0000000, 0.0000000)), ("O", ( 0.0000000, 0.0000000, 1.4500000)), ("O", ( 1.2406484, 0.0000000, 0.9985598)), ("H", ( 0.5158118, 0.8933350, -0.3648961)), # phi=60° ("H", (-1.0316235, 0.0000000, -0.3648961)), # phi=180° (anti-O2) ("H", ( 0.5158118, -0.8933350, -0.3648961)), # phi=300° ], # Glycine (neutral, Cs conformer) — 5 heavy atoms, the second rung # of the speed ladder. Geometry from a tight B3LYP/def2-TZVP # optimisation; only used for wall-clock timing, not parity, so the # exact conformer doesn't need a citation. "glycine": [ ("N", (-1.476, 0.245, 0.066)), ("C", (-0.224, -0.466, -0.093)), ("C", (0.973, 0.450, 0.085)), ("O", (0.901, 1.633, 0.330)), ("O", (2.155, -0.169, -0.043)), ("H", (-1.514, 1.018, -0.595)), ("H", (-2.260, -0.369, -0.112)), ("H", (-0.180, -1.284, 0.638)), ("H", (-0.171, -0.913, -1.094)), ("H", (2.881, 0.456, 0.072)), ], # --- Larger / heavier-element systems for the extended speed # benchmark. All closed-shell. Built from first principles # (Oh / Td / D5h / D2h / all-trans zigzag); not opt-in by default # in SPEED_LADDER_SYSTEMS so `run_speed_benchmark.py` with no flags # stays cheap — pass `--systems naphthalene,ferrocene,...` to # exercise them. # # SF6: octahedral, S-F = 1.561 Å (exp). "SF6": [ ("S", (0.0000, 0.0000, 0.0000)), ("F", (1.5610, 0.0000, 0.0000)), ("F", (-1.5610, 0.0000, 0.0000)), ("F", (0.0000, 1.5610, 0.0000)), ("F", (0.0000, -1.5610, 0.0000)), ("F", (0.0000, 0.0000, 1.5610)), ("F", (0.0000, 0.0000, -1.5610)), ], # CCl4: tetrahedral, C-Cl = 1.766 Å (exp). "CCl4": [ ("C", (0.0000, 0.0000, 0.0000)), ("Cl", (1.0196, 1.0196, 1.0196)), ("Cl", (1.0196, -1.0196, -1.0196)), ("Cl", (-1.0196, 1.0196, -1.0196)), ("Cl", (-1.0196, -1.0196, 1.0196)), ], # Ni(CO)4: tetrahedral Ni(0) d10 (closed shell), Ni-C = 1.838 Å, # C-O = 1.141 Å — the small metal-carbonyl benchmark. "NiCO4": [ ("Ni", (0.0000, 0.0000, 0.0000)), ("C", (1.0612, 1.0612, 1.0612)), ("C", (1.0612, -1.0612, -1.0612)), ("C", (-1.0612, 1.0612, -1.0612)), ("C", (-1.0612, -1.0612, 1.0612)), ("O", (1.7199, 1.7199, 1.7199)), ("O", (1.7199, -1.7199, -1.7199)), ("O", (-1.7199, 1.7199, -1.7199)), ("O", (-1.7199, -1.7199, 1.7199)), ], # Ferrocene Fe(C5H5)2: eclipsed D5h, Fe(II) d6 low-spin (closed # shell). Fe-Cp(centroid) = 1.65 Å, ring C-C = 1.43 Å, # C-H = 1.08 Å. The classic metal-organic-complex benchmark. "ferrocene": [ ("Fe", (0.0000, 0.0000, 0.0000)), ("C", (1.2164, 0.0000, 1.6500)), ("C", (0.3759, 1.1569, 1.6500)), ("C", (-0.9841, 0.7150, 1.6500)), ("C", (-0.9841, -0.7150, 1.6500)), ("C", (0.3759, -1.1569, 1.6500)), ("H", (2.2964, 0.0000, 1.6500)), ("H", (0.7096, 2.1840, 1.6500)), ("H", (-1.8579, 1.3498, 1.6500)), ("H", (-1.8579, -1.3498, 1.6500)), ("H", (0.7096, -2.1840, 1.6500)), ("C", (1.2164, 0.0000, -1.6500)), ("C", (0.3759, 1.1569, -1.6500)), ("C", (-0.9841, 0.7150, -1.6500)), ("C", (-0.9841, -0.7150, -1.6500)), ("C", (0.3759, -1.1569, -1.6500)), ("H", (2.2964, 0.0000, -1.6500)), ("H", (0.7096, 2.1840, -1.6500)), ("H", (-1.8579, 1.3498, -1.6500)), ("H", (-1.8579, -1.3498, -1.6500)), ("H", (0.7096, -2.1840, -1.6500)), ], # Naphthalene: D2h, two fused benzene rings sharing the C-C edge # along the short axis (y). Ring C-C ~1.42 Å, C-H ~1.08 Å. Common # ~20-atom aromatic benchmark. "naphthalene": [ ("C", (0.0000, 0.7080, 0.0000)), # bridgehead, no H ("C", (0.0000, -0.7080, 0.0000)), # bridgehead, no H ("C", (1.2440, 1.4050, 0.0000)), ("C", (2.4370, 0.7160, 0.0000)), ("C", (2.4370, -0.7160, 0.0000)), ("C", (1.2440, -1.4050, 0.0000)), ("C", (-1.2440, 1.4050, 0.0000)), ("C", (-2.4370, 0.7160, 0.0000)), ("C", (-2.4370, -0.7160, 0.0000)), ("C", (-1.2440, -1.4050, 0.0000)), ("H", (1.2440, 2.4870, 0.0000)), ("H", (3.3660, 1.2470, 0.0000)), ("H", (3.3660, -1.2470, 0.0000)), ("H", (1.2440, -2.4870, 0.0000)), ("H", (-1.2440, 2.4870, 0.0000)), ("H", (-3.3660, 1.2470, 0.0000)), ("H", (-3.3660, -1.2470, 0.0000)), ("H", (-1.2440, -2.4870, 0.0000)), ], # n-Decane (C10H22, 32 atoms): all-trans zigzag, C-C 1.54 Å, # C-C-C 113°, C-H 1.09 Å, H-C-H 109.5°. CH2 H atoms in ±y; # terminal CH3 H atoms tetrahedrally distributed. "n-decane": [ ("C", (0.0000, 0.0000, 0.0000)), ("C", (1.2842, 0.0000, 0.8500)), ("C", (2.5684, 0.0000, 0.0000)), ("C", (3.8526, 0.0000, 0.8500)), ("C", (5.1367, 0.0000, 0.0000)), ("C", (6.4209, 0.0000, 0.8500)), ("C", (7.7051, 0.0000, 0.0000)), ("C", (8.9893, 0.0000, 0.8500)), ("C", (10.2735, 0.0000, 0.0000)), ("C", (11.5577, 0.0000, 0.8500)), ("H", (1.2842, 0.8901, 1.4791)), ("H", (1.2842, -0.8901, 1.4791)), ("H", (2.5684, 0.8901, -0.6291)), ("H", (2.5684, -0.8901, -0.6291)), ("H", (3.8526, 0.8901, 1.4791)), ("H", (3.8526, -0.8901, 1.4791)), ("H", (5.1367, 0.8901, -0.6291)), ("H", (5.1367, -0.8901, -0.6291)), ("H", (6.4209, 0.8901, 1.4791)), ("H", (6.4209, -0.8901, 1.4791)), ("H", (7.7051, 0.8901, -0.6291)), ("H", (7.7051, -0.8901, -0.6291)), ("H", (8.9893, 0.8901, 1.4791)), ("H", (8.9893, -0.8901, 1.4791)), ("H", (10.2735, 0.8901, -0.6291)), ("H", (10.2735, -0.8901, -0.6291)), ("H", (-0.3034, 1.0275, -0.2008)), ("H", (-0.7945, -0.5137, 0.5412)), ("H", (0.1877, -0.5137, -0.9428)), ("H", (11.8611, 1.0275, 1.0508)), ("H", (12.3522, -0.5137, 0.3088)), ("H", (11.3699, -0.5137, 1.7928)), ], # n-Hexadecane (C16H34, 50 atoms): same all-trans construction # as n-decane, 6 more CH2 groups. "n-hexadecane": [ ("C", (0.0000, 0.0000, 0.0000)), ("C", (1.2842, 0.0000, 0.8500)), ("C", (2.5684, 0.0000, 0.0000)), ("C", (3.8526, 0.0000, 0.8500)), ("C", (5.1367, 0.0000, 0.0000)), ("C", (6.4209, 0.0000, 0.8500)), ("C", (7.7051, 0.0000, 0.0000)), ("C", (8.9893, 0.0000, 0.8500)), ("C", (10.2735, 0.0000, 0.0000)), ("C", (11.5577, 0.0000, 0.8500)), ("C", (12.8418, 0.0000, 0.0000)), ("C", (14.1260, 0.0000, 0.8500)), ("C", (15.4102, 0.0000, 0.0000)), ("C", (16.6944, 0.0000, 0.8500)), ("C", (17.9786, 0.0000, 0.0000)), ("C", (19.2628, 0.0000, 0.8500)), ("H", (1.2842, 0.8901, 1.4791)), ("H", (1.2842, -0.8901, 1.4791)), ("H", (2.5684, 0.8901, -0.6291)), ("H", (2.5684, -0.8901, -0.6291)), ("H", (3.8526, 0.8901, 1.4791)), ("H", (3.8526, -0.8901, 1.4791)), ("H", (5.1367, 0.8901, -0.6291)), ("H", (5.1367, -0.8901, -0.6291)), ("H", (6.4209, 0.8901, 1.4791)), ("H", (6.4209, -0.8901, 1.4791)), ("H", (7.7051, 0.8901, -0.6291)), ("H", (7.7051, -0.8901, -0.6291)), ("H", (8.9893, 0.8901, 1.4791)), ("H", (8.9893, -0.8901, 1.4791)), ("H", (10.2735, 0.8901, -0.6291)), ("H", (10.2735, -0.8901, -0.6291)), ("H", (11.5577, 0.8901, 1.4791)), ("H", (11.5577, -0.8901, 1.4791)), ("H", (12.8418, 0.8901, -0.6291)), ("H", (12.8418, -0.8901, -0.6291)), ("H", (14.1260, 0.8901, 1.4791)), ("H", (14.1260, -0.8901, 1.4791)), ("H", (15.4102, 0.8901, -0.6291)), ("H", (15.4102, -0.8901, -0.6291)), ("H", (16.6944, 0.8901, 1.4791)), ("H", (16.6944, -0.8901, 1.4791)), ("H", (17.9786, 0.8901, -0.6291)), ("H", (17.9786, -0.8901, -0.6291)), ("H", (-0.3034, 1.0275, -0.2008)), ("H", (-0.7945, -0.5137, 0.5412)), ("H", (0.1877, -0.5137, -0.9428)), ("H", (19.5662, 1.0275, 1.0508)), ("H", (20.0573, -0.5137, 0.3088)), ("H", (19.0750, -0.5137, 1.7928)), ], } ANGSTROM_TO_BOHR = 1.0 / 0.529177210903 # Element symbol -> atomic number. Covers everything in GEOMETRIES: # H/C/N/O (the original molecules + alkanes + ferrocene-C/H), plus # the heavier elements introduced by the extended speed-benchmark # systems: F (SF6), S (SF6), Cl (CCl4), Fe (ferrocene), Ni (Ni(CO)4). _Z = {"H": 1, "C": 6, "N": 7, "O": 8, "F": 9, "S": 16, "Cl": 17, "Fe": 26, "Ni": 28} def geometry_bohr_symbols(system: str) -> List[Tuple[str, List[float]]]: """``GEOMETRIES[system]`` as ``[(symbol, [x, y, z]_bohr), ...]``. The conversion happens **once**, here, with a single constant — and both the ORCA input (fed in Bohr via the ``Bohrs`` keyword) and the vibe-qc side consume the result. That keeps the two codes on *identical* nuclear coordinates, so E_nuc matches to machine precision instead of carrying each code's own Angstrom->Bohr rounding (a ~4e-8 Ha cross-code shift on H2O otherwise). """ return [ (sym, [c * ANGSTROM_TO_BOHR for c in xyz]) for sym, xyz in GEOMETRIES[system] ] def geometry_bohr(system: str) -> List[Tuple[int, List[float]]]: """``GEOMETRIES[system]`` as ``[(Z, [x, y, z]_bohr), ...]`` for vibe-qc.""" return [(_Z[sym], xyz) for sym, xyz in geometry_bohr_symbols(system)] @dataclass(frozen=True) class Cell: """One parity / benchmark cell. Fock-build path: ``df=False, cosx=False`` -> direct 4-index; ``df=True`` -> density-fitted (ORCA RIJK / vibe-qc density_fit); ``cosx=True`` -> RIJCOSX (RI-J + chain-of-spheres K). ``cosx`` implies the DF-J machinery, so a RIJCOSX cell is built with ``df=True`` too. **MP2 cells.** ``method`` may also be ``RMP2`` (closed-shell) or ``UMP2`` (open-shell). For an MP2 cell the ``df`` flag selects the AO->MO transform path: ``df=False`` -> canonical four-index MP2 (vibe-qc) / ``! MP2`` (ORCA); ``df=True`` -> RI-MP2 (vibe-qc ``MP2Options.density_fit`` / ORCA ``! RI-MP2``). ``cosx`` is meaningless for MP2 (no Fock-build K beyond the HF reference) — an MP2 cell with ``cosx=True`` is rejected by :meth:`__post_init__`. """ system: str basis: str method: str # RHF | UHF | RKS-* | UKS-* | RMP2 | UMP2 charge: int = 0 spin: int = 0 # 2S: 0 closed-shell, 1 doublet df: bool = False # density-fitted Fock build (SCF) / RI-MP2 transform cosx: bool = False # RIJCOSX (RI-J + chain-of-spheres K); implies df def __post_init__(self) -> None: if self.cosx and self.is_mp2: raise ValueError( f"Cell({self.cell_id}): cosx is meaningless for an MP2 " f"method — RIJCOSX is an SCF Fock-build path, MP2 has no " f"K build beyond its HF reference." ) @property def is_mp2(self) -> bool: """True for an MP2 / UMP2 correlation cell (vs an SCF cell).""" return self.method.upper() in ("RMP2", "UMP2") @property def cell_id(self) -> str: """Filesystem-safe slug — also the cache-JSON basename. MP2 cells take a ``__RI`` suffix on the DF path (RI-MP2); SCF cells keep the existing ``__DF`` / ``__RIJCOSX`` suffixes. """ if self.is_mp2: suffix = "__RI" if self.df else "" else: suffix = "__RIJCOSX" if self.cosx else ("__DF" if self.df else "") return f"{self.system}__{self.basis}__{self.method}{suffix}" @property def is_open_shell(self) -> bool: return (self.spin != 0 or self.method.upper().startswith("U")) # --- parity matrix ------------------------------------------------------ # Mirrors test_parity_hf_dft.py's _PARITY_CASES diagonal. ORCA covers # three Fock-build paths: direct (4-index), DF (RIJK = RI on J and K), # and RIJCOSX (RI-J + seminumerical chain-of-spheres K). The # BRIEF.md § Scope-1 like-with-like doctrine: every parity cell runs # the same algorithm on both sides — vibe-qc-RIJCOSX is compared # against ORCA-RIJCOSX, not against ORCA's canonical-K-with-DF-J. # vibe-qc's COSX was implemented to match ORCA 6.1.1, so the ORCA-COSX # cells certify implementation consistency between the two # independently-written codes (PySCF has no turnkey RIJCOSX driver, # so the RIJCOSX axis is ORCA-only on the cross-code matrix). PARITY_CELLS: List[Cell] = [ # --- direct Fock build, closed-shell --- Cell("H2O", "def2-svp", "RHF"), Cell("H2O", "def2-svp", "RKS-PBE"), Cell("H2O", "def2-svp", "RKS-B3LYP"), Cell("H2O", "def2-svp", "RKS-PBE0"), Cell("H2O", "def2-tzvp", "RHF"), Cell("H2O", "def2-tzvp", "RKS-PBE"), Cell("H2CO", "def2-svp", "RHF"), Cell("H2CO", "def2-svp", "RKS-B3LYP"), Cell("H2CO", "def2-tzvp", "RKS-PBE0"), # --- direct Fock build, open-shell (OH radical, doublet) --- Cell("OH", "def2-svp", "UHF", spin=1), Cell("OH", "def2-svp", "UKS-PBE", spin=1), # --- density-fitted (RIJK) Fock build --- # RIJK (RI on J *and* K) needs HF exchange, so the DF cells are HF / # hybrid only. A pure-GGA DF cell would have to use ORCA's RIJ # (RI-J only) with a different aux-basis slot — ORCA's RIJK + # def2/JK errors out for a pure functional ("no AuxJ basis set"). # DF-J is exercised by every DF cell anyway; the pure-GGA-DF variant # adds no Fock-build coverage the three cells below don't already # give, so it is intentionally absent. Cell("H2O", "def2-svp", "RHF", df=True), Cell("H2O", "def2-svp", "RKS-B3LYP", df=True), Cell("H2CO", "def2-svp", "RKS-PBE0", df=True), Cell("OH", "def2-svp", "UHF", spin=1, df=True), # --- RIJCOSX Fock build (RI-J + seminumerical chain-of-spheres K) --- # RIJCOSX is K-only-cosx (RI for J, COSX for K) so it requires HF # exchange — HF / hybrid only, no pure-GGA cell (where the K piece # would be empty anyway and the cell would collapse to plain RI-J). # Cells mirror the DF set so each row certifies the same SCF flavour # under all three Fock-build paths (direct, DF, RIJCOSX). All cells # set ``df=True`` because the cosx machinery rides on top of DF-J. Cell("H2O", "def2-svp", "RHF", df=True, cosx=True), Cell("H2O", "def2-svp", "RKS-B3LYP", df=True, cosx=True), Cell("H2CO", "def2-svp", "RKS-PBE0", df=True, cosx=True), Cell("OH", "def2-svp", "UHF", spin=1, df=True, cosx=True), # --- multi-heavy-atom open-shell def2-tzvp (BRIEF.md § Scope-2) --- # Direct + DF coverage on the three open-shell radicals + OH at # def2-tzvp. NO·/def2-tzvp/UHF excluded — vibe-qc's UHF stalls on # this 2Π case where PySCF converges in 19 iter (drop-box # § Regression finding #2). UHF rows on the other systems route # through KDIIS + gtol=1e-6 inside vibeqc_compare's _HARD_OPEN_SHELL # path; UKS rows likewise. The pure-GGA-DF caveat (ORCA RIJK needs # HF exchange) means no UKS-PBE DF cells — only UHF, UKS-B3LYP get # the DF row. Cell("NO", "def2-tzvp", "UKS-PBE", spin=1), Cell("NO", "def2-tzvp", "UKS-B3LYP", spin=1), Cell("CN", "def2-tzvp", "UHF", spin=1), Cell("CN", "def2-tzvp", "UKS-PBE", spin=1), Cell("CN", "def2-tzvp", "UKS-B3LYP", spin=1), Cell("CH3OO", "def2-tzvp", "UHF", spin=1), Cell("CH3OO", "def2-tzvp", "UKS-PBE", spin=1), Cell("CH3OO", "def2-tzvp", "UKS-B3LYP", spin=1), Cell("OH", "def2-tzvp", "UHF", spin=1), Cell("OH", "def2-tzvp", "UKS-PBE", spin=1), Cell("OH", "def2-tzvp", "UKS-B3LYP", spin=1), # def2-tzvp + DF (HF / hybrid only) Cell("NO", "def2-tzvp", "UKS-B3LYP", spin=1, df=True), Cell("CN", "def2-tzvp", "UHF", spin=1, df=True), Cell("CN", "def2-tzvp", "UKS-B3LYP", spin=1, df=True), Cell("CH3OO", "def2-tzvp", "UHF", spin=1, df=True), Cell("CH3OO", "def2-tzvp", "UKS-B3LYP", spin=1, df=True), Cell("OH", "def2-tzvp", "UHF", spin=1, df=True), Cell("OH", "def2-tzvp", "UKS-B3LYP", spin=1, df=True), # --- MP2 / RI-MP2 correlation cells --- # vibe-qc-MP2 vs ORCA-MP2, like-with-like: canonical four-index # MP2 on both sides (df=False) and RI-MP2 on both sides (df=True). # All-electron - vibe-qc explicitly sets n_frozen_core=0 and ORCA is # forced NoFrozenCore (see orca_input.py). Closed-shell RMP2 + open-shell # UMP2 (OH· doublet). cc-pVDZ / cc-pVTZ exercise the f-shell-free # and f-shell correlation regimes; def2-svp keeps a row on the # basis family the rest of the matrix uses. Cell("H2O", "def2-svp", "RMP2"), Cell("H2O", "cc-pvdz", "RMP2"), Cell("H2O", "cc-pvtz", "RMP2", df=True), Cell("H2CO", "def2-svp", "RMP2"), Cell("OH", "def2-svp", "UMP2", spin=1), Cell("OH", "cc-pvtz", "UMP2", spin=1, df=True), ] # The single cell the original investigation says to run first, end-to-end, to nail # the submit -> fetch -> parse -> compare loop before parametrising. FIRST_CELL = Cell("H2O", "def2-svp", "RHF") # --- speed benchmark ---------------------------------------------------- # the original investigation: vibe-qc should be at least as fast as ORCA for the same # method + basis. Sweep system size — a single small molecule is # misleading (setup overhead dominates) — and sweep the Fock-build path, # because that is where the gap lives. The maintainer's own same-box # measurement (2026-05-14, recorded in benchmarks/orca_vs_vibeqc_speed.md) found vibe-qc # ~18x slower than ORCA at HF/def2-TZVP/RIJCOSX on a 15-atom molecule; # the ladder below brackets that with a small system (H2O) and a # medium one (glycine, 10 atoms / 5 heavy), at direct and RIJCOSX. SPEED_LADDER_SYSTEMS: List[str] = ["H2O", "glycine"] # Extended ladder — opt-in via `--systems ` on # run_speed_benchmark.py. Spans 5-50 atoms across organic, aromatic, # row-3 heavier elements, and metal-organic chemistry. NOT the default # because the upper end (n-hexadecane B3LYP/def2-TZVP/RIJCOSX, # ferrocene B3LYP/def2-TZVP/RIJCOSX) takes hours per cell. # # SF6 7 atoms, S+F (row-3), Oh -- small + heavier # CCl4 5 atoms, C+Cl (row-3), Td -- small + heavier # NiCO4 9 atoms, Ni(0) d10, Td -- metal carbonyl # naphthalene 18 atoms, C+H (flat PAH), D2h -- ~20-atom aromatic # ferrocene 21 atoms, Fe(II) d6 LS, eclipsed D5h -- metal-organic # n-decane 32 atoms, C+H (all-trans alkane) -- ~30-atom organic # n-hexadecane 50 atoms, C+H (all-trans alkane) -- 50-atom organic EXTENDED_SPEED_LADDER_SYSTEMS: List[str] = [ "SF6", "CCl4", "NiCO4", "naphthalene", "ferrocene", "n-decane", "n-hexadecane", ] SPEED_LADDER_BASES: List[str] = ["def2-svp", "def2-tzvp"] # RHF (no K shortcut) and RKS-B3LYP (the hybrid where RIJCOSX is the # whole point) span the interesting Fock-build behaviour. SPEED_LADDER_METHODS: List[str] = ["RHF", "RKS-B3LYP"] # Fock-build paths to time, per (system, basis, method): "direct" # (each code's exact 4-index default for these sizes) and "rijcosx" # (RI-J + chain-of-spheres K — the regime where the gap shows up). SPEED_LADDER_PATHS: List[str] = ["direct", "rijcosx"] def speed_cell(system: str, basis: str, method: str, path: str) -> Cell: """Build a benchmark :class:`Cell` for one ladder point. ``path`` is ``"direct"`` (4-index), ``"df"`` (RIJK), or ``"rijcosx"`` (RI-J + chain-of-spheres K). """ if path == "direct": return Cell(system=system, basis=basis, method=method) if path == "df": return Cell(system=system, basis=basis, method=method, df=True) if path == "rijcosx": return Cell(system=system, basis=basis, method=method, df=True, cosx=True) raise ValueError(f"speed_cell: unknown path {path!r}")