Property prediction 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 they do: how hard an alloy is, what light a crystal absorbs, how much gas a porous solid can hold. Property prediction means building a model that takes a description of a substance and returns a number for one of these properties. The description can be a list of which elements are present and in what proportion, a set of averaged properties of those elements, or the full atomic structure: atom types, coordinates and the repeating unit of the crystal. Models used include regression methods such as Gaussian processes and support vector regression, tree ensembles, and neural networks that read structures directly.

Prediction also attaches to things other than the final property. In the examples, a model predicts a temperature at which a synthesis recipe should be run, drawn from the precursors chosen, the target composition and computed chemical driving forces. Another rewrites established equations for how a solder alloy deforms as a neural network, with the equation's terms as neurons and its unknown coefficients as the weights to be fitted. Elsewhere a large language model reads sentences from published papers and pulls out reported property values, with further prompts checking whether each extracted value really is the property claimed, and for a material of the right kind.

Why researchers use it

The honest way to get a property is to make the substance and measure it, or to simulate it from the physics of electrons and atoms. Both are slow and costly. The space of possible materials is far larger than either can cover: an alloy with four or five adjustable elements already has more compositions than a laboratory can test. A fitted model costs almost nothing to evaluate, so it can be run over thousands of candidates to decide which few are worth making. In the examples, predicted values with uncertainty attached are used to choose the next alloy to synthesise, closing a loop between model and laboratory.

Prediction also stands in for scattered or missing data. Property values exist in the published literature but as prose, not tables, so models are used to read them out. Where simulation is affordable but not endless, simulated values can train a model that then generalises to structures nobody simulated. And where a physical equation is already trusted, fitting its coefficients to measurements lets the model inherit the equation's form rather than learn the shape of the physics from scratch.

In the record

Limits

Fitted models work where their training data lies. Asked about an element, a structure type or a temperature range absent from that data, they return a confident number with no warning. Descriptions built from averaged elemental properties discard the arrangement of atoms, so two very different structures can look identical to the model. Labels inherited from simulation carry the simulation's own errors, and labels read from papers carry whatever the papers reported, including values measured on impure or doped samples. Uncertainty estimates help, but they are themselves modelled, and can be too narrow exactly where the model is weakest.

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

Related

• Property prediction: the idea, across every field
• AI in materials and computational chemistry

Every article