We are a group of graduate and undergraduate students at the University of Colorado Boulder, investigating the use of machine learning in experimental physics.
This is partly an experiment in applying machine learning to experimental physics, and partly an experiment in open academic collaboration. Follow our journey in real time!
what?
As part of his thesis, Adam Green was studying the interaction of topological defects in 2D systems. In the lab, we can create quasi-2D fluids by drawing a liquid crystal film over an aperture. We then inject energy into the system (read: hit it really hard), which creates little ‘kinks’ in the fluid, which we can study using polarized-light microscopy. To put it shortly, the interaction of these kinks is interesting to some people (read: physicists).
why machine learning though
This is what our experimental data looks like.
Previous efforts have required the hand-tagging of each defect. Once that is done, statistics can be run on the collection. However, because this is such an arduous process, it severely limits how much data can be analyzed.
Also, I didn’t want to spend months squinting over a screen, clicking at pixels.
So, this seemed like a natural job for machine learning.
but wait…
If you are familiar with machine learning, you know that you have to feed your little robot with training data– ie. data that is tagged.
So have we really won here? We still have to go into experimental data and hand tag it to generate training data.
Well…yes and no. Of course, at the end of the day, we do need some hand-tagged experimental images to verify that our process is giving good results, but because the physical system is fairly simple, we may be able to simulate ‘fake’ experiments on the computer and generate training data.
2D tilted liquid crystal films can be described by the 2D XY model1, which has a rich history in physics. We will discuss the 2D XY model in a future post, but for now, just know that there are methods to numerically solve the 2D XY model. We can use these methods to generate ‘fake’ experiments, which we can then use as training data. At least, that’s the hope. You can read more about the XY model here
-
Not that well though, as the XY model completely neglects hydrodynamics and the liquid-crystal film is, well, liquidy. ↩