The performance of CCSD(T) for the calculation of dipole moments in diatomics

Can a “high-accuracy” method be trusted for every molecular property?

CCSD(T) is routinely used as a reference for electronic-structure calculations and for benchmarking less expensive methods. Much of that reputation has been established through energies, equilibrium structures and related molecular properties. But experiments often depend on different observables. A molecular dipole moment probes the distribution of electronic charge and therefore tests a different aspect of the calculated wavefunction.

The central question was not simply whether CCSD(T) can calculate dipole moments, but whether accuracy demonstrated for one class of properties can be assumed to carry over to another.

Putting high-level theory directly against precision measurements

The study calculated equilibrium structures, vibrational frequencies and dipole moments for 32 diatomic molecules spanning main-group and transition-metal chemistry. The theoretical results were compared directly with carefully selected experimental dipole moments, most derived from high-resolution spectroscopic measurements with small reported uncertainties.

Multiple basis-set families, vibrational averaging and relativistic corrections were examined so that disagreement with experiment could be separated, where possible, from straightforward numerical or methodological approximations.

Accurate overall — but not uniformly accurate

For most molecules, particularly those containing only main-group elements, CCSD(T) reproduced the experimental dipole moments well, with errors typically below 0.15 D. The picture became less uniform for several transition-metal-containing molecules, where substantially larger deviations appeared.

These cases are important precisely because CCSD(T) performs well overall: they identify where a general expectation of high accuracy stops being sufficient and property-specific validation becomes necessary.

Good structures do not guarantee good electron distributions

One of the clearest results was that errors in predicted dipole moments did not track errors in equilibrium bond lengths or vibrational frequencies. Some molecules had accurately predicted structures but noticeably incorrect dipole moments.

The implication is broader than this particular dataset. A calculation that reproduces energies or geometries accurately has not necessarily demonstrated the same accuracy for properties that depend directly on the electron distribution. Computational methods therefore need to be benchmarked against the specific observables for which they will be used.

When theory and experiment disagree

The study examined whether the largest discrepancies could be attributed to familiar sources such as relativistic effects or multireference character. Relativistic corrections were important in specific cases — most notably CuF — but were generally too small to explain the remaining outliers.

AgI provided a particularly instructive example. The CCSD(T) result disagreed strongly with one high-precision experimental value but was close to an earlier measurement from the same experimental group. Rather than assigning the discrepancy automatically to either theory or experiment, the comparison identifies the system as one where both sides warrant renewed scrutiny.

In this sense, a benchmark does more than assign an error bar to a computational method: sufficiently precise predictions can also expose where the experimental or theoretical picture is not yet internally consistent.

Choosing accuracy where it matters

The calculations also showed that computational cost and predictive accuracy do not increase in a simple one-to-one way. For most molecules, the comparatively economical def2-QZVPP basis delivered dipole moments with accuracy similar to the substantially larger aug-cc-pwCVQZ basis. Diffuse functions were most important for particular metal-containing systems and at smaller basis-set sizes.

These results provide practical guidance for extending high-level dipole calculations beyond small benchmark molecules while retaining a quantitatively tested connection to experiment.

Why this matters for predictive theory

Precision experiments increasingly rely on theoretical predictions before every relevant property can be measured. The usefulness of those predictions therefore depends not only on choosing a sophisticated method, but on knowing which observables that method has actually been shown to predict accurately.

This benchmark illustrates a general principle of predictive theory: accuracy should be established against the quantity that matters to the experiment, and disagreement should remain visible rather than being hidden behind the reputation of the computational method. Theory and measurement can then be used together to determine where a microscopic model is reliable, where it requires refinement and where new measurements may be needed.