Gradient-boosted trees in materials and computational chemistry

How it is used

Materials and computational chemistry try to connect what a substance is made of and how it is arranged to what it does: how well it conducts, what colour of light it absorbs, which crystal form it takes, how hot it must be heated to form at all. A decision tree answers such a question by asking a chain of yes-or-no questions about numbers describing the material. Boosting fits many shallow trees in sequence, each one correcting the errors left by those before it, and adds their answers together. The inputs are usually a table: one row per material or per device, with columns holding composition, measured or computed properties, and processing settings.

In this field the tables are typically built first, by other means. Values are pulled out of published papers or taken from large computed databases, cleaned, and turned into numerical descriptors using standard formulae and software toolkits. The trees then sit at the end of that pipeline and predict one quantity from the rest: a property, a class of crystal structure, or a condition for making the material. The examples here show both halves of that arrangement, the assembly and filtering of the table, and the fitted predictor that reads it.

Why researchers use it

Boosted trees suit the shape of the data this field produces. Records are few, columns are many and mixed in kind, some values are missing, and the relationships are not smooth: a small change in composition can switch a material from one phase to another. Trees handle sharp thresholds naturally, do not care whether a column is measured in degrees or electron volts, and need little tuning to work reasonably. They stand in for a physical calculation that would be far slower, or for one that nobody can write down. Where a first-principles simulation would take hours per candidate, a fitted model answers in moments, so many candidates can be ranked.

They also stand in for experience. Deciding what temperature to heat a mixture to, or which family of structure a new alloy will adopt, is normally a judgement made from the literature by someone who knows it well. A model fitted to that same literature turns the judgement into something that can be applied to thousands of untried compositions without a person in the loop, which matters when the prediction feeds an automated experiment rather than a reader.

In the record

Limits

A fitted tree knows only the range it was shown. Asked about a composition or temperature outside that range, it returns a value near the edge of what it has seen, with no warning. Literature tables carry the field's habits: materials that failed are rarely published, so the model learns from successes. Trees give no physical account of why, so a good prediction cannot be checked by reasoning. The descriptors matter more than the method, and a quantity left out of the table cannot be recovered by fitting harder.

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

Related

• Gradient-boosted trees: the idea, across every field
• AI in materials and computational chemistry

Every article