Simulation surrogate in materials and computational chemistry

How it is used

Materials science and computational chemistry try to work out what a substance will do from the arrangement of its atoms: how strong it is, what it absorbs, how it conducts heat or light. The usual tool is a simulation built from physics, most often density functional theory, a way of computing the behaviour of electrons from first principles. It is accurate but slow, so each calculation covers only a small piece of matter for a short span of time. A simulation surrogate is a model trained on the answers such a calculation gives, which then predicts those answers directly for new cases. The physics is run a limited number of times; the surrogate repeats its verdict cheaply.

In the research described here the surrogate takes several forms. Some models learn the energies and forces between atoms, so that the motion of a material can be followed step by step without recomputing the electrons each time. Others take a structure as input, parsed into atom types, coordinates and lattice information, and predict a measured property straight away, such as how much gas a porous framework will take up. Others again learn an intermediate physical quantity, for instance the distribution of electronic charge at an electrode surface, or a correction that extends results from one condition to another. Training data comes from the expensive simulations themselves.

Why researchers use it

The reason is arithmetic. Interesting behaviour in materials often appears only in large systems, over long times, or across many candidate compounds, and the cost of first-principles simulation rises steeply with all three. A surrogate breaks the link between accuracy and cost: once trained, it answers in a fraction of the time, so a search that was impossible becomes a matter of running the model many times over. It stands in for the simulation, which in turn stands in for an experiment.

There is a second reason, which is coverage. A surrogate trained on well-chosen examples can interpolate to conditions that were never computed, such as a different temperature, composition or applied voltage, where running the physics afresh for each would be prohibitive. Some of the work here builds its training set in a loop: the model is used to explore, the cases where it is least confident are picked out, those are computed properly, and the model is retrained. This keeps the expensive calculations aimed at the places where they are most needed.

In the record

Limits

A surrogate is only as good as what it was trained on, and it gives no warning when it strays outside that range: predictions remain fluent while quietly becoming wrong. Structures far from the training set, unusual compositions, extreme pressures or temperatures are the common failure points. Errors can also accumulate, since a small mistake in a predicted force shifts the next step of a simulation. And the surrogate inherits the limits of the physics that taught it, so an approximation in the underlying calculation is reproduced faithfully rather than corrected.

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

Related

• Simulation surrogate: the idea, across every field
• AI in materials and computational chemistry

Every article