Today, I’m attending the launch of the Engineering and Physical Science Research Council’s network on Artificial and Augmented Intelligence for Automated Investigation for Science Discovery. Some great talks already, one of particular note was by Gábor Csányi who has done some fascinating work on training machine learning algorithms to learn atomistic force fields for use in molecular dynamics simulations. In MD, we usually represent atoms, or even groups of atoms, as a single object, so that we don’t worry about the electrons and the nucleus. Newton’s laws of motion are then used to predict how the atoms move in response to other atoms. The tricky part is to figure out the potential that dictates how atoms interact with each other. This is determined by quantum mechanical principles and requires us to approximate the best guess that a full quantum mechanical calculation that includes all the electrons into some simpler function. A favourite, is the Lennard-Jones potential which has a short range repulsion and a long range attraction. In many cases though summarising a complex quantum mechanical prediction into such a potential may be too crude.
Since in a molecular dynamics simulation we don’t care too much about the potential, an alternative is to use something more flexible than a Lennard-Jones, which is where the Gaussian Processes that I’ve been discussing come into play. Essentially Gábor and his team train GPs on quantum mechanical data to find a GP representation of the potential between atoms. The really nice aspect of their work is that they use physics to constrain the possible predictions that the GP can make. One example is the requirement that the simulations are rotationally invariant, so that the properties of the substance do not depend on how we look at the sample.