Candidate generation in materials and computational chemistry

How it is used

Materials science and computational chemistry try to find substances with wanted behaviour: a solar cell that converts more light, an alloy that resists corrosion, a fuel that burns cleanly. The number of possible compositions and structures is vast, so the hard part is deciding what to try next. Candidate generation is the step where a model proposes those options. In the examples, that means turning a chemical description into a fixed list of numbers a model can read, including one-hot codes (a column per element, marked present or absent), stoichiometric ratios, space group and Wyckoff information for crystal symmetry, SMILES text strings for molecules, or computed descriptors of composition and microstructure.

Those encodings then feed models that rank, complete or score proposals. One example drew targets from calculated phase-stability data, used a learned encoding to find literature materials made in similar ways, borrowed their precursors, filled in missing elements with a model trained to guess masked entries, and predicted a synthesis temperature with a tree-based regressor. Others compress a space of device recipes into a few dimensions so it can be mapped, concatenate learned molecular fingerprints into a single vector for a mixture, or generate microstructures by parameterised code and screen their descriptors against simulated mechanical properties inside a closed loop.

Why researchers use it

Researchers reach for these models because the alternative is slow. Making and measuring a material takes days or weeks; a detailed simulation of its behaviour takes computer time that grows quickly with size. Chemists have long relied on intuition and analogy to published work, but that reaches only as far as one person's reading. A model trained on databases of compositions, recipes or calculated properties stands in for that intuition at a larger scale, and it stands in for the expensive step that would otherwise judge each option: a cheap predicted property, or a measure of similarity to something already made, decides which candidates are worth the real cost.

In the record

Limits

A generator can only recombine what its data contains, and chemical databases record what people chose to publish, mostly things that worked. Proposals therefore crowd around known chemistry, and odd but promising options go unseen. Encodings drop information: a composition vector cannot tell apart two materials with the same formula, and a descriptor set is only as good as the formulas behind it. Predicted properties carry errors that the proposal step hides. A candidate that scores well may be unstable, impossible to make, or made only under conditions the model never saw.

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

Related

• Candidate generation: the idea, across every field
• AI in materials and computational chemistry

Every article