Particle Physics is in the fortunate position to be able to collect enormous amounts of data using sophisticated experiments. This data is then used to learn as much about the nature of the subatomic world as possible, hoping to find cracks in our theoretical framework. This is a highly non-trivial task. As modern data science has proven in many other fields to be beneficial in data analysis and statistical inference, my research focusses on tackling the problems in particle physics with modern machine learning techniques.
Optimisation in high-dimensional parameter spaces without known derivative information is a notoriously costly endeavour. There are many different algorithms available, that each approach this problem in a different way. One can however wonder: which of these algorithms works best? This study performs a systematic comparison between a large collection of optimisation algorithms on a diverse set of test functions, including a 12-dimensional stand-in for a particle physics likelihood function. Although the optimisation performance of the algorithms depends on the function to be optimised, we were able to draw some general conclusions based on our experiments.
This study investigates the use of Flow models as stand-in for particle physics event generators. Specifically we investigated whether or not these flow models can be trained on negative weight events, which could e.g. be the computational result of used matching schemes. Turns out the answer is: yes. This is very promissing, as the trained flow models can be used to sample new unweighted events from the distribution formed by the weighted events.
Sampling high dimensional parameter spaces can take a long amount of time. However, if your goal is to find e.g. a boundary in this space, not every part of this parameter space is equally interesting. Typically you are more interested in the region close to the boundary than far away from it. Using machine learning it is possible to spend as little time on the uninteresting part. This technique is called active learning. In this work we show the application of it in the exploration of model spaces and show that it can speed up our searches.
The first use of a machine learning algorithm to accelerate and generalise high-dimensional exclusion limits. We used a RandomForest algorithm to encode both 8TeV and 13TeV ATLAS exclusion limits in a 19-dimensional supersymmetric model. Where a conventional exclusion determination could take up to hours, the trained algorithm is able to do this for any model point within a fraction of a second.
The trained algorithm is called SUSY-AI. It can be downloaded from http://www.susy-ai.org , where also an online version is available.
I believe it is an honerable part of the job to be able to explain the ins and outs of physics (or science in general) to the general public and to be active in a community as a scientist.
I maintain a YouTube channel on which I explain the fundamentals of particle physics through animations and analogies. You can find this channel here.
Apart from the outreach to the general public, i am also active in teaching at Radboud University. Already having a qualification to teach in the first few years of high school, i am now actively working towards a qualitication to teach at any university and university of applied sciences in The Netherlands. A non-extensive list of courses i have been involved in includes: