I’ve spent the past week at the wonderful Lorentz Center at the University of Leiden.
It’s great to know that there are still places that really appreciate that science progresses best by providing an environment that encourage discussion and collaboration amongst scientists with different but complementary backgrounds, and, of course, unlimited coffee or tea! I’ve been finding out about state of the art in the applications of machine learning to soft matter. The field is very much in its infancy so lots of exciting innovations are taking place. Most of the early successes are in using ML to learn force fields for molecular dynamics simulations, as illustrated nicely in a recent paper from one of the presenters:
They have figured out how to constrain the ML predictions to be physically realistic, in this case ensuring energy conservation. Such constraints are not only pleasing from a physics viewpoint, they also help with the challenge that the quantity of data we have to work with in physics is often much less than in other applications of machine learning.