Convolutional neural networks in materials and computational chemistry

How it is used

Materials and computational chemistry try to connect what a substance is made of and how its atoms are arranged to what it does: how it conducts, absorbs light, holds a magnetic state or survives being charged and discharged. Much of the evidence arrives as grids of numbers. An electron or X-ray tomogram is a stack of image slices through a sample. A designed optical structure can be drawn as a pixel map. A diffraction measurement is a long row of intensities against spacing. A convolutional network is built for exactly this shape of data: it slides small learned filters across the grid, so a feature is recognised wherever it appears, and stacks those filters so later layers respond to larger arrangements.

In the examples here, the networks do two kinds of job. One is reading measurements. A segmentation network trained on hand-marked slices of a battery cathode traced each particle through a reconstructed volume, giving every particle its own outline and label so the contact between particle and binder could be counted. The other is standing in for a calculation. A network taking pixel images of small optical building blocks returned how much light each one transmits and with what phase shift, and a classifier taking a molecule's three-dimensional structure decided whether that molecule behaves as a single-molecule magnet. Simulated diffraction patterns are also used as a shared test set on which such models are compared.

Why researchers use it

The pull is partly volume. A tomographic reconstruction holds far more particles than anyone will outline by hand, and marking a few slices to train a network is cheaper than marking them all. Pre-trained weights from a large collection of everyday photographs give the network a usable sense of edges and shapes before it ever sees a sample, so fewer annotations are needed. Known physics can be folded in as an extra constraint during training, as particle shape was in the cathode work, which keeps the output closer to what a particle can actually look like.

In the record

materials-chemistryNeural network segments 650 battery particles to measure their detachment after cyclingA segmentation network, pre-trained on photographs and retrained on marked tomographic slices, produced per-particle outlines and identifiers through a thick battery cathode volume.materials-chemistryNeural network predicts nanostructure optics to design a two-colour laser collimatorA network of stacked convolution and pooling layers read pixel images of meta-atom shapes and returned the transmission coefficient, real and imaginary parts, across the band.materials-chemistryDeep learning sorts crystal structures for single-molecule magnet behaviourA classifier trained on literature labels judged, from three-dimensional molecular structure alone, whether a salen-type metal complex behaves as a single-molecule magnet.materials-chemistryFour million simulated X-ray patterns used to test 21 crystal-symmetry classifiersPhysics-based simulation generated labelled diffraction patterns from filtered crystal structures; learned models appear only as the subject being benchmarked on crystal symmetry classification.

Limits

These networks learn from the examples they are shown, so a sample prepared differently, imaged on another instrument or made of an unfamiliar family of compounds can be read wrongly while the output still looks confident. Labels are a weak point: hand-marked slices carry the marker's judgement, and labels gathered from the literature inherit whatever was measured and reported. A network that stands in for a calculation is only as good as the calculation it copied, and it gives no reason for its answer, so results generally need checking against physics or experiment.

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