Atomic and Molecular Clusters

Do universal AI potentials find the shape of atomic clusters?

A cluster is a tiny clump of a few dozen atoms. The same atoms can arrange themselves in many shapes, and the most stable one decides how the cluster behaves, for example how well it speeds up a chemical reaction. Platinum and gold clusters are among the most studied catalysts.

Born from the forecast

Singularity predicted that the “foundation models” of chemistry, universal AI potentials trained on huge databases of crystals, would arrive in the physics of atomic clusters. They are already used to search cluster structures. We asked whether they can be trusted to pick the right one.

How we tested it

We used a public database of 61,689 cluster structures of 55 elements, each computed with quantum-mechanical calculations (DFT). The protocol was fixed before any comparison: 27 elements for development, 28 kept aside for a single confirmation.

What we found

33%
of cases where the most used universal potential picks the right structure
12%
what a random pick among the candidates would get
42%
with a newer model trained on more varied data
80–87%
if its five best candidates are re-checked with DFT

The potentials recognise cluster geometries but cannot resolve the small energy gaps between competing shapes. Platinum and gold, the metals that matter most for catalysis, are among the hardest cases (21% and 29%).

Why it matters

Studies that compute catalytic or optical properties on the single structure chosen by a universal potential may be computing them on the wrong structure two times out of three. The fix is cheap: use the potential as a filter, not as a judge, and re-check a handful of candidates.

Limits

  • We measure how the potentials rank known low-energy structures, not a full global search, which could only make things worse.
  • Two models were tested in the confirmation.