MACE · AIxSci

What it is

Atoms in a solid or liquid push and pull on one another. Working out those forces from the quantum physics of electrons is accurate but slow, so only tiny systems can be studied for very short times. A machine-learned interatomic potential is a model trained on many such quantum calculations to predict the same energies and forces directly from the positions of the atoms. MACE is one such family. It treats a structure as a graph: each atom is a point, nearby atoms are linked, and information passes along those links so that each atom learns about its surroundings. The model is built to respect rotation, so turning a structure around changes its predicted forces in the matching way rather than giving a different answer.

Some MACE models, such as MACE-MP-0, come ready trained on a large library of known compounds and cover much of the periodic table, so they can be applied to a new material without any new training. They can also be fine-tuned, meaning the ready-trained model is adjusted with a smaller set of fresh quantum calculations for the particular system in hand. Once trained, the model is used in molecular dynamics: the forces are evaluated step by step and the atoms are moved accordingly, so the simulation follows how a material behaves over time.

Why scientists use it

The attraction is cost. A quantum calculation of the forces may allow a few hundred atoms for a few trillionths of a second; a learned potential can be thousands of times cheaper, which opens up larger pieces of material and longer stretches of time at close to the same accuracy. That matters when the interesting behaviour is slow or rare: heat spreading through a crystal, a liquid wetting a surface, a bond breaking and reforming. The model stands in for the quantum calculation, not for the experiment.

A ready-trained model that covers most elements also changes how searches are run. Scientists can screen many candidate structures cheaply, relax them towards their lowest-energy shapes, or use the model's predictions as a first guess that a slower and more exact method then checks. The expensive method is kept for the few cases that matter.

In the record

materials-chemistrySimulated water in graphene gaps shows acid's charge hugging the surfaceStood in for the quantum description of graphene and water, so reactive simulations covering bonds breaking and forming could run at longer times and larger sizes.materials-chemistryTwo neural networks predict how heat shrinks the band gap of silver-based crystalsA ready-trained MACE model was fine-tuned on quantum energies, forces and stresses for small supercells with thermal jiggling and bromine or iodine swaps.materials-chemistryPretrained potentials give a structure search a head start in finding atomic arrangementsA ready-trained universal MACE model informed a search for low-energy atomic structures, alongside quantum calculations and a statistical model fitted to their energies and forces.

Limits

A learned potential only knows what it was trained on. Pushed into unfamiliar chemistry, odd pressures or bonding it never saw, it still returns a confident number, and that number can be wrong in ways the simulation will not flag. It inherits the errors of the quantum method used to make its training data, so it cannot be more accurate than that reference. Ready-trained models covering many elements are convenient but often loose on any one system, which is why they are fine-tuned or checked. Nothing here replaces measurement.

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