M3GNet · AIxSci

What it is

A material is a crowd of atoms held in place by the pull and push between them. To know whether an arrangement of atoms is stable, scientists need its energy: roughly, how comfortable the atoms are where they sit. They also need the forces, which say which way each atom would move. These can be calculated from the physics of electrons, but that calculation is slow. M3GNet learns the answer instead. It treats a material as a graph: a set of points, the atoms, joined by lines to their near neighbours, with the distances and the angles between those lines recorded. A neural network, a set of adjustable numbers tuned on examples, passes information along those lines and returns an energy and a force for each atom.

The examples it learns from come from the slow electron-level calculations, run on many thousands of known and candidate materials. Because the model only ever looks at an atom's local surroundings, it can be applied to a material it has never seen, including one with a different number of atoms or a different mix of elements. That is why it is called universal: one trained model covers most of the periodic table rather than a single chemical family. Given a starting arrangement, it can be asked for forces again and again while the atoms are nudged downhill, until they settle into a shape where nothing is pulling hard any more.

Why scientists use it

Speed is the whole point. The electron-level method, density functional theory, is the usual standard for how a material behaves, but a single structure can take hours of computer time, and a search over many compositions and many ways of arranging the atoms needs thousands of structures. A learned model answers in a fraction of a second. That changes what questions can be asked: instead of testing a handful of candidates, a researcher can sweep a whole family.

It is often used as a first pass rather than a replacement. Rough starting guesses are tidied up by the model, which carries the atoms most of the way to their settled positions, and only then is the slow, trusted calculation run on the nearly finished structure. The slow method then has less work to do and still has the final say. The model can also supply an informed starting expectation for a search method that goes on to refine it with real calculations.

In the record

materials-chemistryTesting whether symmetry-aware neural networks can tell apart atomic arrangements in perovskitesPre-relaxed many perovskite supercells with differently arranged cations before slower density functional theory produced the final structures, energies and distance from the convex hull.materials-chemistryPretrained potentials give a structure search a head start in finding atomic arrangementsSupplied a prior expectation for a Bayesian search over atomic structures, whose Gaussian process was fitted to energies and forces from density functional theory.materials-chemistryZero-padding lets one neural network encode crystals with differing element countsContext for encoding inorganic structures as fixed-size inputs built from atomic numbers, stoichiometric ratios, elemental feature vectors, space group and Wyckoff site information.

Limits

The model repeats patterns in the data it was trained on. For chemistry unlike that data, such as odd oxidation states, strongly magnetic materials or atoms squeezed far closer than usual, it can be confidently wrong, and it gives no warning when it is. Learned energy surfaces also tend to come out too smooth, so small energy differences between near-identical arrangements, often the ones that decide which structure wins, may be blurred away. It predicts energies and forces, not electronic or optical behaviour, and it does not replace the slower calculation when an answer must be trusted.

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