Self-learning computers will become an important tool in the search for unknown particles and particle processes. This is the conclusion of a group of researchers from Nikhef and Radboud University, among others, in a recently published article on Arxiv about the search for outliers.
The researchers looked at the performance of four different self-learning systems in detecting deviations from the Standard Model of particle physics. Especially combinations of such systems seem promising, says former Nijmegen PhD student Melissa van Beekveld, recently theorist at Oxford. There are many models and ideas and only limited resources and people to follow them all. Machines can help to select the interesting ideas and look at them further’.
The group around RU researcher Sascha Caron, also Nikhef, has been working for some time on applications of artificial intelligence in particle physics. Among other things, they are looking at usability in the ATLAS experiment, the KM3NeT neutrino telescope, in gravity waves and gamma radiation in cosmic radiation. There are now hundreds of self-learning algorithms for this type of research. Soon a ranking will be published of these algorithms, made by among others Caron’s PhD student Luc Hendriks.
Particle physics is all about predicting particle patterns that arise when protons or other particles collide, for example in the LHC accelerator at CERN in Geneva. The vast majority of collision processes are neatly described by the equations of the Standard Model. That model is not perfect or complete. In which direction it should be extended and how, is not clear. Theorists have countless ideas about this.
Now such ideas and models are tested by researchers of experiments like ATLAS by searching for predicted deviations. Many of these tests are unsuccessful: no deviations from the Standard Model are found and the conclusion is that at least there is no new physics hidden there.
Such zero results have two problems. First, to the outside world particle physics does not seem to be able to discover anything beyond the known particle physics for years. And apart from that image problem there are actually too many ideas about the world beyond standard physics to test them all thoroughly.
Self-learning systems search for outliers in particle processes that no longer fit within the framework of the Standard Model. To do so, they learn what is normal by absorbing the expected patterns in particle collisions. In the test of the four self-learning techniques use was made of artificial data, created with so-called Monte Carlo techniques according to the rules of the Standard Model.
The new study focused on how each system reacts after that training to processes that are unthinkable in the Standard Model, so-called anomalies. Some of the studied techniques turn out to spot such impossibilities easier than others, says Van Beekveld. But the bottom line is that machine learning is a useful tool in the search for new physics. This is definitely the way forward, we think.
Machine learning is a rapidly growing interest in particle physics, and the number of papers in that field has been rising sharply in recent years. But it is unlikely, as Van Beekveld also says, that one specific approach or technique will deliver the breakthrough. Machine learning is a broad concept, something like mathematics. An enormous toolbox from which you ultimately take what you need. As far as that’s concerned, we’re really still at the beginning.