Linear models in materials and computational chemistry

How it is used

Materials science and computational chemistry try to work out why a substance behaves as it does from the arrangement of its atoms: how strongly they bind, how they vibrate, how light is absorbed, which structures form. Much of this is done by solving approximate equations of quantum mechanics on a computer, most often by a method called density functional theory, which gives energies and forces for a given set of atomic positions. A linear model is a fitted equation in which the quantity of interest is a weighted sum of chosen ingredients, and only the weights are learned. Here those ingredients are numbers describing local atomic environments, compositions or lattice occupations.

In the research on this site, linear fits appear inside larger calculations rather than as the headline. Weighted sums of structural descriptors are fitted to quantum-mechanical energies and forces, producing cheap substitutes for the expensive calculation that can then drive simulations of atoms in motion. Similar fits map which lattice sites are occupied onto an energy, so that many candidate arrangements can be ranked quickly. Elsewhere a plain least-squares line converts results computed under one set of conditions to another, standing in for calculations that would otherwise have to be repeated.

Why researchers use it

The pull is cost. A quantum-mechanical calculation of a single structure can take hours of computer time, and a simulation may need millions of such evaluations. A fitted sum of descriptors returns an answer almost instantly, so a small set of carefully chosen calculations can be stretched to cover a much larger search. Because only weights are learned, the fit can be solved directly rather than searched for, which makes refitting cheap enough to repeat many times as new data arrive.

Linear models also offer something more practical than accuracy: a usable sense of their own reach. When a new structure falls outside the range spanned by the fitting data, that distance can be measured, and the examples here use exactly this to decide which structures to compute next and add to the training set. The weights themselves can be read and checked against known chemistry, which matters in a field where a wrong energy quietly produces a plausible but false structure.

In the record

Limits

A weighted sum cannot represent behaviour that is not close to a sum of its chosen parts, so accuracy depends entirely on whether the descriptors capture the chemistry involved. Bond breaking, charge transfer and magnetic effects are often poorly served. Outside the range of the fitting data the predictions degrade, sometimes smoothly and invisibly, and a simulation driven by such a fit can settle into a structure that looks reasonable and is wrong. In the examples the response is iteration: compute more, refit, and check for extrapolation rather than trust the fit alone.

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

Related

• Linear model: the idea, across every field
• AI in materials and computational chemistry

Every article