Predicting dipole moments of diatomic molecules

A property shaped by charge redistribution

Elemental composition specifies which atoms are present in a molecule, but molecular polarity emerges only after their electronic structures interact.

Dipole moments are therefore sensitive to aspects of bonding that are not fully represented by elemental identity alone.

This makes them a useful test of how the required molecular representation changes when the target observable depends more directly on electronic redistribution.

Identifying the relevant physical information

We used Gaussian-process regression to compare different combinations of atomic and molecular descriptors across the 162-molecule dataset.

Atomic quantities including electron affinity and ionization potential carried clear predictive information. The best models, however, also required molecular features associated with the formed molecule.

With these combined descriptors, dipole moments could be predicted with relative errors of approximately 5%.

The comparison between representations was as important as the final prediction accuracy: it showed which types of physical information become necessary for an observable that cannot be described adequately at the composition level alone.

Representation follows the observable

The dipole-moment results complement the spectroscopy models without requiring the same representation.

Several ground-state spectroscopic properties can be learned from highly compact information about the constituent elements. Excited-state spectroscopy requires additional atomic and molecular-state information. Molecular dipole moments require descriptors that carry information about charge redistribution within the bonded molecule.

Together, these results establish a practical hierarchy of molecular representations:

composition → atomic information → molecular information → detailed microscopic description

The useful level of description depends on the target property. A representation that is sufficient for one observable may omit essential physics for another.

Connecting rapid prediction with detailed modelling

For molecules with limited existing data, compact representations can provide an initial property estimate without immediately requiring a full electronic-structure treatment.

Where the observable depends on richer molecular physics, the representation can be expanded accordingly.

This provides a natural progression from rapid molecular-property prediction to more detailed microscopic calculations, with model complexity introduced where the target property actually requires it.