CHGNet · AIxSci

What it is

A material is a pattern of atoms. To work out how it behaves, scientists need its energy, and the forces pulling each atom towards a more comfortable position. Those quantities can be calculated from quantum mechanics, but the calculation is slow. An interatomic potential is a cheaper stand-in: a formula, or in this case a trained model, that reads the positions of the atoms and returns the energy and forces directly. CHGNet learns that mapping from a large set of quantum calculations done in advance.

It does so as a graph neural network. The atoms become points, and nearby atoms are joined by links, so the structure becomes a graph rather than a list of coordinates. The network passes information repeatedly along those links, so each atom builds up a description of its own chemical neighbourhood, and the energy is assembled from those descriptions. The extra ingredient in CHGNet is charge: alongside energy and forces, it predicts how much electronic charge sits on each atom. In many inorganic materials, especially oxides, the same element can carry different charges in different sites, and that changes everything about how it bonds.

Why scientists use it

Scientists use models like this because the honest quantum calculation is too slow to run very often. Searching for a new battery material, or a stable arrangement of atoms in a crystal, can mean assessing thousands or millions of candidate structures, or following the atoms as they jiggle over time. A trained potential makes each assessment fast enough that the search becomes possible, and the expensive calculation is kept for the few candidates that survive.

In the record

materials-chemistryZero-padding lets one neural network encode crystals with differing element countsTurned compositions, space group numbers and Wyckoff positions into a fixed-size numerical description of a crystal, padding shorter entries with zeros so every structure matched in length.materials-chemistryTesting whether symmetry-aware neural networks can tell apart atomic arrangements in perovskitesCubic forty-atom perovskite cells were enumerated across seventy-two elements, pre-relaxed with a trained potential, then settled with quantum calculations to give energies and stability.

Limits

A trained potential only knows the chemistry it was shown. Offered an element, a pressure or a bonding arrangement unlike anything in its training data, it will still return a confident number, and that number may be wrong. Small errors in the forces can accumulate over a long simulation and send the atoms somewhere the real material would never go. The charges it reports are learned regularities, not a description of the electrons themselves. Results that matter are normally rechecked with the slower quantum calculation.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Model family

Graph neural network

Every article

2articles

0structural-biology

2materials-chemistry

0astronomy

aix-00147

Testing whether symmetry-aware neural networks can tell apart atomic arrangements in perovskites

property predictionsimulation surrogate

materials-chemistry · ai-result · arXiv · 2024

aix-00111

Zero-padding lets one neural network encode crystals with differing element counts

candidate generationproperty prediction

materials-chemistry · ai-result · arXiv · 2026