Structure determination in materials and computational chemistry

How it is used

Materials chemistry asks what a substance is made of and how its atoms are arranged, because arrangement decides behaviour: whether a surface speeds up a reaction, whether a crystal conducts, whether a molecule is stable. The arrangement is rarely visible. Researchers infer it, either from indirect measurements such as spectra (the pattern of light a sample absorbs, which depends on how its atoms are bonded) or by calculating the energy of candidate arrangements from the physics of electrons and keeping the lowest. In this work, machine learning sits inside that inference. It groups structural descriptions into manageable categories, learns the relationship between an arrangement and its energy, and stands in for calculations that would otherwise be too slow to run.

The calculations being replaced or shortened are usually density functional theory, a way of computing the energy and forces of a fixed set of atomic positions from quantum mechanics. It is accurate and expensive, so each arrangement tested costs real computer time, and a search over arrangements may need very many. The models used are modest and statistical rather than large: clustering to divide a continuous structural measure into groups, least squares and nonlinear fits to rescale computed spectra to different surface conditions, and Gaussian processes, which predict an energy together with a measure of how uncertain that prediction is. Training data comes from the expensive calculations themselves.

Why researchers use it

The reason is cost. A search for the right structure is a search through an enormous space of possible atomic positions, and the only trustworthy judge of each candidate is a quantum calculation. Researchers reach for a learned model as a cheap stand-in: it guesses which candidates are worth the expense, so the expensive method is spent sparingly. A model that also reports its own uncertainty is useful here, because the search can be steered towards places where the guess is both promising and poorly known.

The same logic applies when comparing a calculation with an experiment. A spectrum can be computed for an idealised arrangement, but the real sample may be crowded, disordered or at a different coverage, and recomputing for those conditions is expensive. A fitted correction can carry the cheap result across to the harder condition. Learned models are also used to supply training signals where none exist: forces and stresses can be invented from simple physical rules, so a model can be taught to respond to a distorted structure without quantum data for every case.

In the record

materials-chemistryNeural networks read infrared spectra to work out how CO sits on platinumClustering grouped a continuous measure of atomic neighbourhood into site classes, and fitted scaling factors rescaled computed infrared frequencies and intensities to higher surface coverage.materials-chemistryPretrained potentials give a structure search a head start in finding atomic arrangementsA Gaussian process trained on energies and forces from randomly generated configurations, using a structural fingerprint, guided a search for low-energy atomic arrangements.materials-chemistryRandom atoms settle into molecules and crystals on a learned energy surfaceEquilibrium structures were randomly displaced and distorted, then given invented forces from a spring and a repulsion, and stresses from an elastic rule, as training targets.

Limits

A learned stand-in is only as good as the calculations it was trained on, and those calculations carry their own approximations, so a confident prediction can be confidently wrong. Training sets built from random or slightly perturbed arrangements cover familiar territory, and models extrapolate poorly to chemistry they have not seen. Invented forces and fitted corrections depend on parameters chosen by hand rather than measured. Agreement between a predicted spectrum and an observed one does not prove the structure is right, since different arrangements can produce similar measurements.

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