Learning diatomic molecular properties from composition information

Molecular information from minimal inputs

When two atoms form a molecule, their electrons reorganize, chemical bonds emerge and new electronic states become available. Molecular properties are therefore not simple combinations of the properties of isolated atoms.

Equilibrium bond lengths, vibrational frequencies and binding energies are normally obtained from spectroscopy or electronic-structure calculations that describe the molecule itself.

At the same time, spectroscopic measurements accumulated over many decades reveal systematic relationships between these quantities. Diatomic molecules provide a useful setting for testing how general those relationships are: they contain only two atoms, but span very different electronic structures and bonding mechanisms.

Ground-state spectroscopy across the periodic table

Based on our database of experimental spectroscopic properties of heteronuclear diatomic molecules, we used Gaussian-process regression to model three fundamental ground-state quantities:

• equilibrium bond length,
• vibrational frequency,
• binding energy.

The representation was deliberately compact. Information related to the groups and periods of the two constituent elements already captured a substantial part of the structure in the data.

Across chemically diverse molecules, several ground-state properties could be predicted with relative errors below approximately 5%.

These predictions require neither a molecular geometry nor a detailed electronic representation. They show that a substantial amount of information about some molecular observables is already encoded in elemental composition.

Excited states require a richer description

The same elemental composition can support different electronic states with different equilibrium structures and vibrational spectra.

We extended the analysis to the first excited electronic state and modelled its equilibrium distance and vibrational frequency.

Composition remained informative, but additional atomic properties and spectroscopic information from the molecular ground state were needed. With these additional inputs, prediction errors remained below approximately 11%.

The comparison shows that the information content of a representation depends on the property and electronic state being studied. Ground-state observables can sometimes be strongly constrained by composition, while excited-state properties require a more detailed physical description.

Screening poorly explored molecular space

The availability of molecular data is highly uneven.

Some diatomic species have been characterized extensively by experiment and high-level electronic-structure calculations. Others have only sparse measurements or remain largely unexplored, despite the wide range of element combinations available across the periodic table.

This matters because poorly characterized molecules may possess optical, electronic, chemical or spectroscopic properties that have not yet been recognized.

Composition-based models provide a low-cost first layer for exploring this space. They can estimate missing properties, expose broad chemical trends and help identify systems that merit more detailed calculations or experiments.

For the observables studied here, useful predictive accuracy could already be obtained from a very small amount of input information. This makes composition-level prediction a practical screening layer before more expensive microscopic modelling is applied.

Later extension

This research line was later extended to a broader collection of heteronuclear and homonuclear diatomic molecules, with additional descriptor comparisons and tests of transfer between molecular classes in Spectroscopic constants from atomic properties: a machine learning approach, Digital Discovery (2024), DOI 10.1039/D3DD00152K ↗.