DLPNO-CCSD(T) · ORCA User Guide

On this page

Why DLPNO-CCSD(T)
Basic usage
Precision tiers — Loose / Normal / Tight
Complete basis-set extrapolation (CBS)
F12 — fast approach to the basis-set limit
Open shell — UHF reference
Recommended workflow
One step further — multireference (CASSCF · NEVPT2)

Why DLPNO-CCSD(T)

CCSD(T) is the "gold standard" of single-reference quantum chemistry. For typical closed-shell species it reproduces relative energies to within ±1 kcal/mol — the so-called chemical accuracy. The catch is that canonical CCSD(T) scales as O(N⁷), which makes it prohibitively expensive for systems beyond ~20 atoms.

The DLPNO (Domain-based Local Pair Natural Orbital) approximation dramatically reduces this cost. Using local orbitals and domain-based pair natural orbitals it achieves asymptotically linear scaling while recovering about 99.9 % of the canonical CCSD(T) correlation energy. The manual summarises the situation simply: "if DLPNO-MP2 is feasible, DLPNO-CCSD(T) is feasible too." In practice this puts molecules of several hundred atoms within reach.

Basic usage

Add DLPNO-CCSD(T) to the keyword line together with an auxiliary basis:

# Closed-shell single point — a typical precision energy
! DLPNO-CCSD(T) cc-pVTZ cc-pVTZ/C TightSCF

* xyzfile 0 1 mol.xyz

The auxiliary basis (the /C suffix) follows the same convention as RI-MP2. DLPNO-CCSD(T) calculations require RI integrals, so the auxiliary basis is mandatory.

Main basis	Recommended auxiliary basis
`cc-pVDZ`	`cc-pVDZ/C`
`cc-pVTZ`	`cc-pVTZ/C`
`cc-pVQZ`	`cc-pVQZ/C`
`def2-SVP`	`def2-SVP/C`
`def2-TZVP`	`def2-TZVP/C`
`def2-TZVPP`	`def2-TZVPP/C`

Precision tiers — Loose / Normal / Tight

DLPNO cutoff thresholds are consolidated into three tiers:

Keyword	Typical cost	Relative accuracy	Use case
`LoosePNO`	0.5×	~95–98 %	Large-scale screening; quick exploration.
`NormalPNO`	1× (default)	~99.5 %	General research. The default recommendation.
`TightPNO`	~2–3×	~99.9 %	Benchmarks, precision energies prior to publication.

# Benchmark-level accuracy
! DLPNO-CCSD(T) TightPNO cc-pVTZ cc-pVTZ/C TightSCF

Use the same threshold across compared species

When comparing reaction energies or bond-dissociation energies, apply the same PNO threshold to every species so that the errors cancel. Mixing NormalPNO on one and TightPNO on another can introduce ~0.5 kcal/mol of inconsistency.

Complete basis-set extrapolation (CBS)

CCSD(T) is most accurate when extrapolated to the complete basis-set limit. ORCA's Extrapolate keyword automates this:

# Auto-extrapolate DLPNO-CCSD(T) from cc-pVTZ and cc-pVQZ
! DLPNO-CCSD(T) Extrapolate(3/4,cc) AutoAux TightSCF

* xyzfile 0 1 mol.xyz

Syntax: Extrapolate(X/Y,basis) — no space after the comma, or it fails to parse. X and Y are cardinal numbers (2 = DZ, 3 = TZ, 4 = QZ, 5 = 5Z); basis is the family name — cc (Dunning), def2, aug-cc, etc. With DLPNO you must also add AutoAux (it builds the auxiliary basis on the fly).

Input	What ORCA actually does
`Extrapolate(2/3,cc)`	Two calculations (cc-pVDZ and cc-pVTZ), then extrapolate.
`Extrapolate(3/4,cc)`	cc-pVTZ + cc-pVQZ extrapolation. The most common choice.
`Extrapolate(2/3,def2)`	def2-SVP + def2-TZVPP.
`Extrapolate(3,cc)`	Three calculations (cc-pVDZ, cc-pVTZ, cc-pVQZ) + two pairwise extrapolations.

EP2 · EP3 — efficient extrapolation for larger bases

Running CCSD(T) directly with a large basis (e.g. cc-pV5Z) is often too expensive, but MP2 may be feasible. EP2 / EP3 combine "large-basis MP2 + small-basis (CCSD(T) − MP2) correction" to keep accuracy while controlling cost.

# EP2: small basis for CC, large basis for MP2 extrapolation
! DLPNO-CCSD(T) ExtrapolateEP2(2/3,cc) AutoAux TightSCF

F12 — fast approach to the basis-set limit

F12 explicit-correlation methods reach near-CBS accuracy with a smaller basis. DLPNO-CCSD(T)-F12 combines the two ideas:

# DLPNO-CCSD(T)-F12 with approach D — F12 replaces the large-basis effect
! DLPNO-CCSD(T)-F12D cc-pVDZ-F12 cc-pVDZ-F12-CABS cc-pVDZ-F12/C TightSCF

F12 requires its dedicated basis plus a CABS (Complementary Auxiliary Basis Set); don't forget the -F12 and -F12-CABS suffixes. With F12, a single cc-pVDZ-F12 calculation can match the accuracy of a cc-pVQZ / cc-pV5Z CBS extrapolation.

Open shell — UHF reference

For radicals or multiplet systems, use a UHF reference ("UHF-DLPNO-CCSD(T)"):

! UHF DLPNO-CCSD(T) cc-pVTZ cc-pVTZ/C TightSCF

* xyz 0 2          # doublet radical
  O   0.0   0.0   0.000
  H   0.0   0.0   0.969
*

Check for spin contamination

With a UHF reference, always check the SCF ⟨S²⟩ value. The ideal value for a doublet (S = 1/2) is 0.75, but severe spin contamination can push it above 1.0, in which case the subsequent CCSD(T) result is unreliable. If contamination is severe, try an ROHF reference or move to a multi-reference treatment (CASSCF / NEVPT2).

Recommended workflow

DLPNO-CCSD(T) is usually used as a single-point method. Analytical gradients are available but expensive, so the standard "composite" workflow is to optimise with a cheap method (DFT) and refine the energy with CCSD(T) on top.

Optimise + frequencies: B3LYP-D4/def2-TZVP or r2SCAN-3c. Obtain the thermal correction ΔG_therm.
DLPNO-CCSD(T) single point: CBS extrapolation or F12 on the optimised geometry — high-precision electronic energy E_elec^CC.
Solvation correction: if needed, ΔG_solv from (DFT/SMD) − (DFT/gas).
Final free energy: G_final = E_elec^CC + ΔG_therm^DFT + ΔG_solv.

Run this flow as three separate files (not one $new_job input — see the warning below):

# -- step1.inp -- DFT geometry + frequencies (D4 dispersion ON) -> step1.xyz
! B3LYP D4 def2-TZVP RIJCOSX def2/J TightOpt Freq
* xyzfile 0 1 mol.xyz

# -- step2.inp -- DLPNO-CCSD(T)/CBS single point (no dispersion: CC already has it)
! DLPNO-CCSD(T) Extrapolate(3/4,cc) AutoAux TightSCF
* xyzfile 0 1 step1.xyz

# -- step3.inp -- solvation (SMD), on the same optimized geometry
! B3LYP D4 def2-TZVP RIJCOSX def2/J CPCM(water) TightSCF
%cpcm
   SMD true
   SMDSolvent "water"
end
* xyzfile 0 1 step1.xyz

Two places this composite dies (checked on ORCA 6.1.1)

① Do not chain the three steps with $new_job in one input. The D4 dispersion from the DFT optimisation carries into the CC step and triggers a “dispersion + correlated method” error (CCSD(T) already includes dispersion, so ORCA blocks it). Split the steps into files and keep D4 out of the CC step.
② DLPNO + Extrapolate requires AutoAux (otherwise an “auxiliary basis needed” error), and Extrapolate(3/4,cc) must have no space after the comma or it fails to parse.

When reporting reaction equilibria or free-energy barriers, this "DFT geometry + CCSD(T) precision energy + SMD solvation" combination is the most standard recipe in modern quantum chemistry and is routinely accepted by journals.

DLPNO with other methods

The DLPNO approximation is not limited to CCSD(T); it is also available for CCSD, MP2, and B2PLYP (double hybrid). For example, ! DLPNO-B2PLYP D4 def2-TZVP def2-TZVP/C cuts double-hybrid cost substantially — very useful for precision single points on medium-to-large molecules.

One step further — multireference (CASSCF · NEVPT2)

DLPNO-CCSD(T) is the apex of single-reference theory. But where two or more Slater determinants carry comparable weight — molecules midway through bond breaking, non-colinear magnetic states in transition-metal and lanthanide complexes, singlet diradicals — the single-reference assumption breaks down. From there CASSCF, which solves the active space explicitly, is the starting point, with NEVPT2 adding dynamic correlation.

Choosing the active space, state-averaging, NEVPT2, spin-orbit coupling · SINGLE_ANISO, and the metal/ligand fragment merge now live in their own chapter → 13 · CASSCF · Multireference.