Recurrent neural networks in materials and computational chemistry

How it is used

Materials science and computational chemistry study what substances are made of, how their atoms are arranged, and what properties follow from that arrangement. Much of the data comes as ordered sequences: a molecule written as a line of text, a diffraction pattern read across angles, a signal recorded over time. A recurrent network processes such a sequence step by step, keeping an internal state that it updates at each step, so earlier parts of the sequence can influence how later parts are read. Long short-term memory layers, a common variant, add gates that decide what to keep and what to discard, which helps the state survive over longer stretches.

In practice the network is often one branch of a larger model. One study encoded molecules as SMILES strings, a text notation for chemical structure, fed them through stacked long short-term memory layers, and joined the result with separately computed numerical descriptors of the same molecules before predicting combustion properties. Elsewhere, sequence-shaped data is used to test models rather than to build them: a benchmark of simulated X-ray diffraction patterns, each a long vector of intensities, was assembled by physics simulation and labelled by crystal system and space group, with learned models appearing only as the thing being measured.

Why researchers use it

Researchers reach for recurrent networks when the order of the data carries meaning and the useful length varies. A molecule's text notation has no fixed size, and which atoms matter depends on what surrounds them, so a method that must be given a fixed grid of numbers in advance is awkward. The recurrent state stands in for hand-built rules about which parts of a sequence to combine: instead of deciding in advance which fragments of a molecule drive a property, the network learns what to carry forward as it reads.

They also stand in for measurement. Properties such as how a fuel burns are slow and costly to determine in the laboratory, and only a small fraction of possible molecules or mixtures has ever been tested. A model trained on what has been measured gives a cheap estimate for what has not, so candidates can be ranked before anything is synthesised. Because one network can be trained to predict several properties together, it can also express the trade-offs that matter when a material has to satisfy more than one requirement at once.

In the record

materials-chemistryDeep-learning model designs petrol blends from octane and soot targetsEncoded molecules as SMILES text through stacked long short-term memory layers, combined with normalised molecular descriptors, to predict three combustion properties together.materials-chemistryFour million simulated X-ray patterns used to test 21 crystal-symmetry classifiersSupplied simulated diffraction patterns, labelled by crystal system and space group, as a benchmark; the patterns came from physics simulation, not from a learned model.

Limits

A recurrent network reads a molecule's written form, not the molecule, so two notations for the same structure can be read differently, and the model learns the conventions of the text as well as the chemistry. Memory still fades over long sequences. Predictions hold only where training data was dense, and chemical space is mostly empty, so estimates for unusual structures can be confidently wrong. The learned state is hard to inspect, which makes it hard to say why a prediction was made, or to tell a real physical relationship from a quirk of the dataset.

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