Clustering in materials and computational chemistry

How it is used

Materials and computational chemistry study what substances are made of, how their atoms arrange themselves, and how those arrangements set properties such as strength, conductivity or how well a device converts light into electricity. Clustering means letting a computer sort data into groups by similarity, with no labels supplied by a person. In this field it is pointed at measurements and at descriptions of composition. In the examples, diffraction patterns, which record how X-rays scatter off a sample and so reveal its atomic arrangement, are reduced to a small number of recurring components or divided into regions that share a phase. Elsewhere, a continuous measure of how an atom sits among its neighbours is cut into discrete groups.

Why researchers use it

The pull is volume and the absence of labels. An experiment can sweep across a composition or a heat treatment and return a pattern at every point, far more than anyone will interpret by hand, and nobody knows in advance which phases are present. Grouping by similarity supplies a first answer: here are the few distinct things in this data, and here is where each one appears. It also compresses. A composition can be written as a long vector of element counts, mostly zeros, and reduced to a handful of coordinates in which similar chemistries sit close together, so that later models have fewer, better-behaved inputs to work with.

In the record

materials-chemistryA learning algorithm ran a synchrotron beamline to find a phase-change materialSpectra were turned into similarity weights on a composition graph, and diffraction patterns were segmented into phase regions by a graph method encoding physical constraints.materials-chemistryRobot laser-heats thin films while software picks each next heating conditionDiffraction patterns along an annealed stripe were reduced by non-negative matrix factorisation to four components, with amorphous ones then discarded.materials-chemistryStatistical model of 5,441 perovskite solar cells predicts efficiency and suggests recipesPerovskite compositions encoded as element vectors were reduced to four components by locally linear embedding, and the resulting space inspected visually.materials-chemistryNeural networks read infrared spectra to work out how CO sits on platinumK-means grouped a nearly continuous coordination measure into discrete classes, so site structure could be written as a probability distribution.

Limits

Groups found this way are statistical, not physical. The number of components or clusters is usually set by the researcher, and a different choice gives a different picture; some components correspond to nothing chemically meaningful and have to be identified and thrown out by hand. Cutting a continuous quantity into bins discards real variation. A reduced space is convenient but hard to interpret, since its coordinates have no chemical meaning. And similarity in a measurement does not guarantee sameness of structure, so results need checking against physical knowledge.

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