Large language models in materials and computational chemistry

How it is used

Materials and computational chemistry study what substances are made of, how they behave and how to make them, using a mixture of experiment, simulation and the written record of past work. A large language model is a program trained on enormous amounts of text to continue and transform writing. In this field it is usually pointed at sentences from papers, asked to pick out a material, the property being reported, its value and its units, and then asked follow-up questions about each item: is this really the property it claims to be, is the sample pure, what is the chemical formula, was the number measured or computed. It can also be asked to generate chemistry rather than read it: which starting chemicals to combine, and at what temperatures.

In the examples here the models are given instructions in plain English and, sometimes, a handful of worked cases to imitate, with no retraining. Their answers are then put through conventional chemistry software to be standardised, and scored against held-back data: exact matches for proposed starting chemicals, average error for predicted temperatures. One study queried several different models through a single interface and retried when an answer could not be parsed. So the model is a component in a pipeline, not the whole of it, and its output is checked by other tools.

Why researchers use it

Most of what is known about materials sits in prose: a value in a sentence, with the conditions and caveats spread across the paragraph around it. Older extraction software matches patterns and so cannot easily tell a band gap from another quantity with similar wording, or a pure crystal from a doped one, or a measured number from a calculated one. Those judgements need the sense of the sentence, and that is what a text model offers.

On the making side, deciding which starting chemicals and which temperatures will yield a target compound is a judgement built from many previous recipes. A model that has read widely can propose a ranked set of candidates without being told how many ingredients to use, which is awkward to encode in rules. In both roles the model stands in for a trained person reading a paper or recalling a recipe, at a scale that no person could cover.

In the record

materials-chemistryGPT-4 reads band gap values from paper sentences to train better predictorsPulled materials, properties, values and units from paper sentences, then answered follow-up prompts on whether each entry was a genuine band gap, a pure crystal, or calculated.materials-chemistryLanguage models suggest ingredients and firing temperatures for inorganic materials recipesProposed ranked sets of starting chemicals for inorganic syntheses, and predicted heating temperatures as structured data, scored against held-out text-mined recipes.

Limits

A text model has no grasp of chemistry, only of how chemists write, so it can return a well-formed formula that does not exist, a value lifted from the wrong clause, or a plausible recipe that no furnace would realise. It states wrong answers as confidently as right ones, which makes silent errors the main risk when it is run over many sentences at once. Its training text overlaps the literature being mined, so apparent accuracy may be recall. Everything it produces needs standardising and checking against held-back data or experiment.

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