vibe-qcDocumentation

DLPNO: ORCA feature parity, measured efficiency and extensions¶

Status assessed on 2026-11-03 against main 1a0ab4fc, the local primary literature and ORCA 6.1 manual content identifying itself as 6.1.1. #508 owns this programme; #591 records scientific qualification. The priority is ORCA DLPNO feature parity first, then the best measured complete-method time and memory at matched accuracy, then new capabilities for molecules and solids. This plan adds no release gate or release date.

The molecular energy engine is useful and has bounded accepted evidence. Full ORCA feature parity and production scaling are not established. Shared Fock, MANTA, symmetry, MPI and GPU infrastructure must serve the missing methods as well as accelerate accepted ones. Infrastructure alone does not supply a missing equation or scientific acceptance.

Capability and evidence matrix¶

The reference inventory is the official ORCA 6.1 MP2, local correlation and NEVPT2 chapters, supplemented by sections 5.12/5.13 for excited states and 5.36 for LED. Every comparison must pin its executable version and resolved settings; a manual filename or matching keyword is insufficient.

Capability	Current implementation/evidence	Remaining parity obligation
Restricted DLPNO-MP2/CCSD	PAO/PNO domains, pair equations, physical occupied coupling, raw canonical limits; 17-molecule CC table and #707 strict water/HF controls	Complete versioned recipes, domain/pair ladders, broader chemistry and nonminimal basis/fit qualification (#852/#730)
Restricted local T0	Per-triple TNO correction and nonzero/full-space controls	Matched ORCA T0 settings and complete scaling (#591/#190)
Default restricted `t1`	Occupied rotation restores canonical triples from local amplitudes at canonical cost	True reduced-scaling local iterative T1 (#851); do not label current compatibility behavior as that algorithm
Independent-UHF MP2/CCSD/triples	Separate spin spaces, frozen/minority/closed limits; experimental dense iterative triples option	Broad radical accuracy and PAO locality; ORCA high-spin ROHF/QRO/NEV-PNO construction is a separate missing formulation (#590)
Loose/Normal/Tight, SCS/SOS, core policies	Published CC cutoff triples and explicit companions; UMP2 spin scaling exists	MP2-specific recipes, TightPNO coupled-LMP2 semantics, complete spin/core/weak-pair policy and faithful refusals (#852)
Local CISD/QCISD	Canonical variants exist; local CCSD does not supply them	Variant-specific residuals, corrections and high-spin scope (#860)
MP2 derivatives and field response	Generic derivative machinery exists; no qualified local analytic family	Restricted analytic gradient/relaxed density, then electric/magnetic second derivatives (#854)
CCSD density	Canonical Lambda algebra is available as a reference	Local Lambda and closed-shell orbital-unrelaxed density (#853); this is not CCSD(T) density or analytic CC nuclear gradients
F12	Orbital/CABS basis data exist	MP2/CC F12 C/D operators, projectors and CABS singles; reference-specific qualification (#855)
LED and multilevel	Fragment/EDA foundations and pair data are reusable	Published LED components/preparation and multilevel pair policy (#856/#859)
Complete-PNO/CBS	Explicit Altun extrapolation primitive with negative controls	Matched endpoint provenance, policy qualification and separate basis limit (#859/#730)
Excited states	Canonical response foundations are reusable	Back-transformed PNO EOM/STEOM and genuine local STEOM are different targets; local STEOM is closed-shell in the reference manual (#857)
Multireference	Canonical CAS/NEVPT2/RDM foundations exist	Local fully internally contracted NEVPT2 and dependent F12 (#858)
Source/Fock/MANTA	Shared PAO/PNO/TNO solve sites and explicit evaluated/selected-RI bridges have bounded acceptance	Typed native evaluated-state admission and broader source/reference rows (#580/#581)
Symmetry/MPI/GPU	Shared contracts and selected primitives exist	Complete DLPNO consumers and measured accuracy/cost (#589/#807/#803/#190)
Solids	Real Gamma/BvK adapters reuse the molecular engine; experimental translation MP2 and internal CC residual machinery exist	Separate reference/localization, iterative CC/triples/spin and physical convergence obligations (#572/#582/#586/#587/#588)

ORCA’s open-shell local CC uses common high-spin ROHF/QRO orbitals and SOMO-aware pair spaces. It is not a general broken-symmetry independent-UHF algorithm. Its UHF-MP2 uses independent spin spaces, so the MP2 and CC reference contracts must also remain distinct. The numerical equality of one full-space UHF calculation does not qualify the QRO local approximation.

The accepted #707 controls and #730 fit separation remain valid at their exact settings. In the latter, six new fixed-orbital water full-space controls have maximum raw CCSD error 8.03e-14 Ha. Frozen-core cc-pVDZ-JKFIT correlation differs from the direct calculation by -89.0412 microhartree, while the direct/archived-ORCA total difference is -0.0245467 microhartree. Thus internal exactness, auxiliary-fit error and external accuracy cannot be represented by a single PASS label.

Work order and independent pickups¶

Stage	Deliverable	Tracker and first bounded pickup
A0	Honest convergence and method identity	#830 skips triples on failed CC amplitudes; #845 fixes public correlation convergence; #852 records versioned recipes
A1	Ground-state energy feature parity	#851 restricted iterative local T1; #590 high-spin QRO/NEV-PNO pair construction; #730 one admitted cc-pVTZ control; #860 full-space variant residual
A2	Density and derivatives	#853 fixed-domain local adjoint/density; #854 restricted full-domain MP2 Lagrangian/gradient, then local response
A3	Remaining released local families	#855 geminal/CABS operators; #856 closed-shell component reconstruction; #857 full-space excited sigma action; #858 one FIC excitation class; #859 matched endpoint validator
B	Complete-method efficiency at equal quality	Existing #580/#581/#589/#807/#803/#804/#818/#819/#190, one measured consumer at a time
C	New scientific reach	Later periodic CC/spins, Lambda triples, higher excitations and other literature-supported extensions with separate acceptance

These are capability dependencies, not a requirement to finish every umbrella before independent development. Correctness-oriented source work and profiling can proceed in parallel; performance promotion waits for a qualified comparison. F12 operator work can begin while local T1 develops; its CC consumer needs the relevant accepted CC route. MP2 gradients need their own Lagrangian, not completion of CC Lambda. Local MR F12 depends on its specific accepted MR and F12 operands. Missing full-text equations are an explicit dependency for the corresponding implementation, not a reason to halt unrelated ready work.

Each tracker issue specifies the first bounded deliverable, code entry points, mathematical assumptions, negative controls and acceptance. Workers claim exact files and scope, preserve active owners, and land source on main with CI skipped. Source delivery, affected-method vq acceptance and performance qualification are separate records. The same worker may perform development and validation; no extra reviewer or coordinator gate is added.

Theory that implementation must preserve¶

For the closed-shell convention of Riplinger and Neese, 2013, Sec. II, Eqs. 23-24,

[ \widetilde T_{ab}^{ij}=\frac{4T_{ab}^{ij}-2T_{ba}^{ij}}{1+\delta_{ij}}, \qquad D^{ij}=\widetilde T^{ij}(T^{ij})^\dagger+ (\widetilde T^{ij})^\dagger T^{ij}. ]

Changing this density’s normalization changes the meaning of its occupation cutoff. Pair multiplicities, screened/weak-pair energies and the untruncated-minus-PNO correction remain explicit. The DOI of normalized PAOs in SparseMaps II, Sec. II, Eq. 1 is a density-overlap integral, not a distance radius. The two-stage preparation and sparse maps are part of the algorithmic cost model.

For pair bases U, transport uses S_ij,kl = U_ij^H S_AO U_kl. Coupled LMP2/CC residuals require occupied-Fock couplings and these interpair overlaps. A diagonal denominator alone is a guess or declared approximation. Distributed ghosts must therefore carry the same amplitude iteration, source and coordinate frame as their target residual.

Guo 2018, Eqs. 2-4 distinguishes iterative triples with occupied-Fock/inter-TNO transport from diagonal T0, and specifies the large-minus-small-space T0 correction. In ORCA nomenclature (T) means T0 and (T1) means the improved iterative treatment. This T1 name is unrelated to the singles-amplitude Hamiltonian transformation in Jiang 2024.

Saitow 2017, Sec. II A/B, Eqs. 3-12 uses the high-spin Dyall/SOMO construction for common-orbital PNOs; Guo 2020 supplies its triples companion. Neither should be cited as proof that arbitrary independent-UHF pair spaces implement the same method.

The local adjoint satisfies stationarity of L = E_CC(T) + Lambda.R(T), including the adjoint of every pair transport. It supplies an orbital-unrelaxed density. Additional orbital/localization constraints are required for relaxed properties. For MP2, Pinski 2019, Eqs. 17-24 differentiates SCF, localization, kept/discarded PNO and semicanonical-amplitude constraints, as well as one-electron, overlap/Pulay, localization, AO and RI metric integrals. A derivative with frozen local subspaces is a useful diagnostic but is not the complete analytic gradient. Discrete cutoff crossings need declared selection/response policies. Electric/magnetic second derivatives use additional controls, as in Stoychev 2021.

Kumar 2020, Sec. II A/B separates ordinary CCSD, conventional triples, CABS singles and F12 corrections. The geminal, strong-orthogonality projector and orbital-plus-CABS resolution are additional physical operands. Approximation C/D and omitted contractions must be identified explicitly; a scalar basis correction is not F12.

Schneider 2016 LED requires fragment/reference preparation and component reconstruction beyond pair energies. Dutta 2016 and the ORCA excited-state chapters distinguish back-transformed and local algorithms. Guo 2016 local FIC-NEVPT2 requires contracted metrics/RDMs and class-specific amplitude systems; a strongly-contracted denominator is a different approximation. Its 2023 F12 extension has additional projector and CABS obligations.

Unified implementation and performance contracts¶

Fock/eigensolver: consume #472/#580 OperatorFrame, physical FockBuild and shared solvers. Preserve F C=S C epsilon, global metric orthogonality, stationary occupied density and independent canonical frozen boundaries. Never use DIIS/shifted F as the physical correlation operator or treat a pair-density/CC Jacobian as a Hermitian Fock problem.
MANTA: consume #581’s actual selected factors, metric whitener, rank/projector, finite support and source identity. Reuse #617/#638 operators and transforms. A native MANTA SCF receipt alone does not establish a same-source correlation fit. No second integral engine.
Symmetry: #589 first checks molecular factor/projector covariance, then reduces equivalent pair work with correct multiplicities. Nuclear symmetry alone cannot authorize reduction of a broken-symmetry density, frozen partition or finite-support producer. Periodic sewing/unfolding is a separate capability.
Tensor and massive parallel: #807 owns ragged pair/triple consumers; #804/#820 supply ownership, bounded transfers and retirement. Fixed target-residual ownership and epoch-matched ghosts precede distributed iteration. Preserve #801’s active periodic panel work. No all-pair gather, global-rank padding or duplicate scheduler.
GPU: #803/#639 provide optional operations, leases and precision contracts. Qualify a measured dominant contraction, then a complete changed calculation. The current Apple G1m DF-shaped operation failed its pilot accuracy gate; no production DLPNO GPU result follows from other successful pilot operations. Native FP64, complex conjugation, explicit mixed-precision correction and device-specific evidence matter.
Cost: reuse #819 measurement records and #190 scaling ownership. Report cold/warm complete time, preparation/iteration/triples/output, peak owned/replicated/device storage, communication, throughput and resource-hours at identical scientific settings. Preserve the current allocation blocker; do not launch a competing C40/C60 campaign.

The objective of being fastest requires named competitors, hardware, workloads, accuracy and reproducible measurements. A faster integral tile, smaller fitting basis, looser PNO cutoff or unconverged iteration does not establish a complete-method speed advantage.

Qualification and literature feedback¶

Keep four independent axes in #591: raw same-finite-source canonical exactness; local/preset approximation; fit/basis/physical convergence; external matched accuracy. Molecular raw CCSD 1e-8 Ha, triples 1e-9 Ha, residual 1e-11 and energy change 1e-13 Ha remain unchanged for their applicable controls. The periodic raw full-space gate is separately 2e-9 Ha/cell. Derivatives, response, RDM and component tolerances require predeclared units and convergence studies, not reuse of an energy tolerance for a different observable. Record all domain/PNO/screened corrections apart from the raw comparison.

The full feature matrix must include compatible solvent, scalar-relativistic, ECP/core, restart/output and numerical-derivative workflows. Test both supported combinations and explicit refusals. Pin output labels to the actual formulation, reference and converged state. Old operands are useful controls but do not qualify changed consumers.

Library follow-up should acquire missing foundational full texts/SI for density (10.1063/1.4962369), SparseMaps IV/V (10.1063/1.4945444, 10.1063/1.4979993), early MP2 gradients (10.1063/1.5011204), multilevel (10.1021/acs.jctc.7b00260) and open-shell LED (10.1021/acs.jctc.8b01145). Search recent papers and errata through the assessment date, including integral-direct MP2-F12 (10.1021/acs.jctc.5c02184), without treating uninspected abstracts as working equations. Return source/SI paths, derivations, assumptions, counterexamples, benchmark data/licensing and cost evidence to refine these issues. Current library entries for higher excitations, Lambda triples and periodic work remain research-extension evidence until scoped.

Solid-state continuation and boundaries¶

The 2025 periodic DLPNO-MP2 BvK and megacell papers describe distinct TURBOMOLE development formulations. They are expansion targets, not released ORCA periodic parity. CRYSTAL23 remains the PBC mean-field reference authority; correlation needs an appropriate independent method or justified numerical limit.

#591 note 58711 records a bounded chi-LiH support endpoint with numerical and memory qualification; note 60382 owns its ongoing independent external audit. It is not yet an external-accuracy result. Preserve that owner and the historical failed assessment outputs. Neither they nor the accepted translation MP2 controls establish periodic canonical CCSD(T) PASS. The real BvK-supercell adapter remains separate from translation-native CCSD iteration/triples/spins under #586/#587/#588. Broader periodic method development remains deferred in the current work sequence and outside #557/QC0.18. The existing experimental labels and immutable runtime pins remain part of the contract.