Data Analysis in High-Energy Physics as a Differentiable Program

Author
Affiliation
Nathan Simpson

Lund University

Published

February 3, 2023

Preface

“I didn’t have time to write a short letter thesis, so I wrote a long one instead.” ― Mark Twain (adapted)

Many curious things can stem from direct messages on Twitter1. You can count the contents of this thesis as one of them.

In October of 2019, I got off the tram from the Blandonnet stop in Geneva, arriving bewildered to CERN for a 6 month placement. I had no idea what I was going to be doing there, since my assigned supervisor and I had not communicated at any point leading up to this visit. I just knew that I was doing this placement as part of my position, which was funded as part of a Marie Curie International Training Network (they let you do placements), and that my supervisor was in some way connected to the team working on the ROOT software suite.

Now, ROOT is sometimes talked about with a negative tone in high-energy physics circles due to it’s steep learning curve and monolithic software design. I will endeavour to do no such thing here, as this software was a revolution in computing for its time, and continues to power the majority of data analysis done at the Large Hadron Collider. However, this is specialist software, and it is very much the case that the typical undergraduate/masters student will likely have no training on how to use it, which can significantly hinder the ability to play with your code and understand what’s going on. Moreover, many courses at university will be teaching students to have proficiency with more standard scientific analysis tools, such as the scientific Python ecosystem. Being one of these people, I was keen to explore to what extent I could utilize these tools that I had some familiarity with in order to do my physics work. This led to me tweeting out the below, a mere two months before my placement:

I enthusiastically followed up on Lukas’ offer to collaborate, and before I knew it, I was sat at a table in CERN’s main restaurant (R1 if you’re a cool kid) with both Lukas and my would-be supervisor (who would not be my supervisor), ready to start a project on building something that had never been done before, but using tools that I was familiar with. That project became neos, and started the development of a new paradigm for how we view optimizing our data analysis pipelines, and the cornerstone of this thesis – differentiable programming. Totally different from just a ROOT-less analysis, but I also did a bit of that in the end anyway.

As such, if you were expecting a typical HEP thesis, you may be surprised at what lies ahead, hopefully in a positive matter. We will still traverse the lands of particle physics, but will be stepping around many puddles that most dive into, and instead spend a bit of extra time in the marshlands of probability and statistics, which inform almost all of the work done in my PhD from a practical and motivational standpoint. From there, we’ll foray into the fields of gradients, including how we calculate them, why we care, and what they’re used for. There, we will meet the buzzwords that you may have come here for in the first place: machine learning and differentiable programming. Armed with these fundamentals, we will tackle my application of them to problems in collider physics, including optimizing summary statistics while being aware of uncertainty, interpolating between new physics models, and searching for new particles produced in association with a Higgs boson! (Maybe there’s puppies too. I don’t know. But don’t you want to find out?)

This whole thesis was created from markdown and executable code, thanks to the Quarto framework.

  1. Citation needed, but not provided.↩︎