Graph neural networks in materials and computational chemistry

How it is used

Materials and computational chemistry try to work out what a substance will do from the arrangement of its atoms: how strong it is, what colour it appears, how it conducts electricity, how it reacts. The usual tool is quantum mechanics, solved on a computer, which is accurate but slow. A graph neural network treats a molecule or crystal as a graph: each atom is a point, and nearby atoms are joined by links. The network repeatedly sends short messages along those links, so each atom gathers information about its surroundings, and then predicts a quantity such as the total energy or the forces pulling on each atom.

In this field the commonest use is as an interatomic potential, meaning a fast stand-in for the quantum calculation that gives energies and forces for any arrangement of atoms. Those forces can then be used to run molecular dynamics, a simulation that moves the atoms step by step and shows how the material behaves over time. Some such models are trained to cover many chemical elements at once, so they can be pointed at unfamiliar compounds. Others start from a general trained model and are adjusted on a smaller set of quantum calculations for one particular system. Graph-style descriptions of structure are also used as input features for predicting properties directly.

Why researchers use it

The appeal is cost. Quantum calculations of the kind used to generate training data are expensive enough that only small systems and short timescales are affordable, which rules out many questions: what happens at a surface in contact with water, how a solid behaves as it warms, how a defect moves. A trained network answers in a tiny fraction of the time, so the same simulation can be run for longer and on larger boxes of atoms.

The graph itself is why this kind of network suits the problem. Chemistry has no fixed list of positions to fill: molecules differ in size, crystals differ in composition, and the same substance looks the same if you rotate it or renumber its atoms. A graph with messages along the links handles all of that naturally, and the message passing can be built to respect rotation, so predicted forces point the right way. It also means a model can be trained on one set of structures and applied to others that were never seen, which is what discovery work needs.

In the record

materials-chemistryA single neural network stands in for quantum calculations across 45 elementsA graph neural network potential with tensor message passing was fitted to energies, forces and charges from molecular, crystal and public datasets, with the calculation settings supplied as an extra input.materials-chemistrySimulated water in graphene gaps shows acid's charge hugging the surfaceAn equivariant message-passing potential stood in for the quantum energy surface, so reactive dynamics including bond breaking could run beyond the reach of direct quantum simulation.materials-chemistryTwo neural networks predict how heat shrinks the band gap of silver-based crystalsA pre-trained general-purpose interatomic potential was fine-tuned on quantum energies, forces and stresses for supercells with thermal displacements and substituted halogen atoms.materials-chemistryZero-padding lets one neural network encode crystals with differing element countsComposition, symmetry group and atomic site data were packed into a fixed-size input combining pre-computed elemental feature vectors with padding for structures of differing length.

Limits

A trained potential is only as good as the quantum calculations behind it, and it inherits their errors as well as their accuracy. Pushed to arrangements unlike anything in training, it can give confidently wrong forces, and nothing in the prediction announces that it is out of its depth. Training sets drawn from different sources often use inconsistent settings, which must be reconciled somehow. The networks also give little insight into why they answer as they do, and a smooth, fast surrogate can quietly miss rare events, charged or magnetic states, and long-range effects that reach beyond the graph's links.

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