~/aixsci
weekly updates

fields/materials-chemistry/what the AI was for

Attribution in materials and computational chemistry

How researchers in materials and computational chemistry use machine learning to work out what accounts for a material's behaviour: which inputs, variables or structural features produced an outcome, and how much each mattered.

How it is used

Materials and computational chemistry study why a substance behaves as it does: why one compound conducts heat poorly, why another holds a magnetic state. Attribution asks a particular question about this. Not what the property is, but what brought it about, and in what proportion. The usual approach is to fit a model that maps measurable or calculated quantities onto an outcome, then inspect the model to see which quantities carried the weight. The inputs may be numbers from physics simulations, such as how a material conducts electricity or heat at a given temperature. They may instead be the arrangement of atoms in three dimensions. In both cases the model is a means of apportioning credit among candidate causes.

Two patterns appear in the research here. In one, quantities computed from first principles (calculations that start from the behaviour of electrons rather than from measurements) are fed in as separate inputs, and the fitted relationship shows how each contributes to the figure of merit being sought. In the other, a model reads molecular structure alone and sorts molecules into those that show a behaviour and those that do not. The second kind attributes an outcome to shape and arrangement, by establishing that structure is sufficient to tell the two classes apart.

Why researchers use it

Attribution in this field is hard by hand because the candidate causes are entangled. In a thermoelectric material, the quantities that govern performance pull against one another: improving one tends to spoil another. Writing down which matters most, at which temperature, is not something a single equation settles. Machine learning stands in for a systematic sweep of the possibilities. Given a body of calculated or reported outcomes, it finds the combination of inputs that tracks the result, and the fitted model then says something about their relative weight.

It also stands in for intuition about structure. Chemists have working rules about which arrangements of atoms should produce a given magnetic or electronic behaviour, but the rules are partial and were built from a small number of cases. A model trained on reported results tests whether structure alone accounts for the behaviour, without the researcher having to name in advance which structural feature is responsible. Where labels come from the published literature, this turns scattered experimental reports into a single statement about what distinguishes one class from the other.

In the record

Limits

A fitted model shows which inputs track an outcome, not which cause it. Where inputs are themselves related, as conductivity and thermal conductivity are, weight can be assigned to the wrong one. Inputs taken from simulations inherit the simulation's approximations, so an attribution can reflect the method rather than the material. Labels gathered from published work carry the literature's biases: what was measured, under which conditions, and what went unreported. A model that separates two classes from structure has shown structure suffices to tell them apart, not which structural feature matters.

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

Related

Every article