When I was young, I calculated higher-order cross sections in supersymmetry for a code package called Prospino2. In those good old days I also worked on proper simulations of physics beyond the Standard Model with a Monte Carlo generator called MadGraph.

Even though my enthusiasm for supersymmetry searches has taken hit after hit during the LHC running, my group still takes pride in simulating things properly. QCD is just too much fun to use short cuts.

During the early LHC years the Higgs boson has transitioned from the most exciting discovery to a search path for physics beyond the Standard Model. The way to search for such effects is through the Standard Model effective field theory. Coordinated with some field theory work we established a global Higgs and gauge boson fit through our SFitter tool – run by happy French-German coalition.

In SFitter we are having great fun with frequentist and Bayesian statistics applied to LHC data. Our main goal is to get the uncertainties right, even those theory uncertainties which are not even defined properly. And since we are at is, we recycle all this technical know-how for global top-sector analyses, including higher-order predictions, or global dark matter analyses.

Talking about statistics, together with friends at NYU we developed a way to compute and understand where information about signal processes at the LHC actually come from; which regions of phase space we should study, and what we can expect from different production and decay channels.

This application of information geometry is a great example how to apply proper maths to LHC physics without following of the usual path of formal physics. So if you are interested in learning what the Neyman-Pearson lemma or the Cramer-Rao bound have to do with the LHC, or if you want to know why physicists like to call a score an optimal observable, just drop me an email.

For years, we have been working, first on old, and now on modern machine learning for LHC physics. It all started with the question how we can identify boosted top quarks using QCD jet algorithms, where our HEPTopTagger really made a difference. We then upgraded it to DeepTop, based on convolutional networks, and DeepTopLoLa, our first boost-symmetry-aware network. There we were able to measure the Minkowski metric with error bars in MadGraph events – and to get that published.

As for most people, ML-classification was just the entrance to scientific AI. technical playground. Everything we always wanted to do seems possible now. Going back to my interest in statistics, I am now a big fan of Bayesian or statistical ML. Then, there are generative networks entering the world of Monte Carlo simulations, and conditional generative networks for unfolding. And all of this is just the start of an infinite number of technical projects re-shaping theoretical particle physics with modern machine learning…