Autoencoders in materials and computational chemistry

How it is used

Materials science and computational chemistry try to connect what a substance is made of, and how its atoms are arranged, to what it does: how hard it is, how it conducts, whether it forms at all. An autoencoder is a pair of networks. The encoder squeezes a description of a material into a short list of numbers, and the decoder tries to rebuild the original from that list alone. Because nothing else is available to the decoder, the short list has to retain whatever matters. In this field the input is usually a built description of a compound. One example converts composition, symmetry group and atomic site information into a single fixed-size vector, padding shorter entries so that structures of different complexity can sit side by side.

Variants add structure to that compressed space. A variational autoencoder learns a spread of plausible values rather than one point, so the space can be sampled and new candidate compositions proposed. Attention, a mechanism that lets the network weigh some parts of the input more heavily than others, can be added to the encoder. One example works on bulk metallic glasses, alloys that solidify without forming crystals, encoding each composition as a normalised vector across many elements together with a measurement condition, and relating these to recorded hardness.

Why researchers use it

The difficulty is input, not output. A molecule or crystal is not naturally a row of numbers. Compositions vary in length, symmetry is categorical, atomic sites come as nested lists, and two samples of the same material may be described at different levels of detail. Most learning methods want a fixed-size numerical input. An autoencoder earns its place by learning that fixed-size form from the data itself, with reconstruction as the only instruction, rather than relying on a chemist's chosen set of descriptors.

The compressed space is also useful in its own right. Similar materials tend to land near each other, which allows a property model to be fitted on the short list instead of the raw description, and allows unexplored regions to be sampled for candidates worth making. This matters because measured data is scarce: properties such as hardness are often assembled from published experiments, one composition at a time. Compression lets a small labelled set go further, since much of the learning uses unlabelled descriptions, and physical knowledge can be written into how the input is encoded.

In the record

materials-chemistryZero-padding lets one neural network encode crystals with differing element countsCombined one-hot atomic numbers, stoichiometric ratios, elemental feature vectors, space group and Wyckoff site data into one fixed-size vector, zero-padding shorter entries within each batch.materials-chemistryNeural network designs five hard metallic glasses, all confirmed by experimentEncoded literature metallic glass compositions as normalised vectors across 56 elements, with indentation load as a scalar, for variational learning of Vickers hardness.

Limits

Reconstruction is not understanding. An autoencoder learns to rebuild whatever it was shown, so its compressed space inherits the biases of the input encoding and of the compositions that happen to have been published. Hand-built representations can discard information that matters, and padding or normalising to a fixed size can make unlike materials look comparable. Sampled candidates may be chemically implausible or impossible to synthesise, since nothing in the training enforces stability. The compressed numbers have no physical meaning, which makes it hard to see why a prediction came out as it did.

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