Random forests in materials and computational chemistry

How it is used

A decision tree is a chain of yes/no questions on measured quantities that ends in a prediction. A random forest grows many such trees, each on a different random sample of the data and of the available inputs, then averages their answers. In this field the inputs are usually a table: the fractions of each element in a material, quantities derived from them by established formulae, or the settings of a laboratory procedure. The output is a property that is slow to measure or compute, such as which crystal phase forms, an electronic band gap, how fast a metal corrodes, or the colour of a treated oil.

Much of the work sits before the forest itself. Compositions are turned into numerical descriptors, odd values removed, gaps filled in, and the numbers rescaled so that none dominates. Forests are then often one entry in a line-up of methods, fitted alongside support vector machines, gradient-boosted trees and neural networks on the same split of the data, with the best performer kept. They are also used the other way round: because each split in a tree records how much it improved the prediction, a forest can rank which descriptors carry the information, a question in itself.

Why researchers use it

Datasets here are small and awkward. They are stitched together from published tables, from simulations that each take hours, or from experiments run a few dozen times. A forest copes with a few hundred rows, mixed units, missing entries and inputs that are strongly related to one another, and it needs little adjustment to work tolerably. Averaging many trees also damps the tendency of a single tree to memorise its training set.

In the record

Limits

A forest only averages the examples it saw, so it predicts poorly for compositions or conditions outside the range of its training table, and tends to flatten towards the middle rather than flagging that it is lost. It carries no physical or chemical law, so a prediction can be internally plausible and still impossible. Literature-assembled data inherits whatever the literature over-reported. Importance rankings are unreliable when descriptors are derived from the same composition and so move together, since the forest can credit either one.

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

Related

• Random forest: the idea, across every field
• AI in materials and computational chemistry

Every article