Hi there! đď¸
My name is Liam, and I recently joined the ranks at Datameister as the âPhysics Meisterâ. To fulfill this role, I have brought with me a unique set of experiences in the wonderful world of high-energy particle physics experiments.
While it's commonly understood that physics contributes significantly to AI, the reverse is also true. In future blogs, I plan to delve deeper into this fascinating intersection between my previous work and the innovative world of AI that we are tackling at Datameister. But for now, I'd like to start by introducing myself and sharing the journey that led me here.
Humble beginnings
My career in physics started when I first enrolled as a student of Physics and Astronomy at Ghent University in 2013. As I delved deeper into my studies, I became fascinated by particle physics, which explores the tiniest components of our universe. I decided to follow this passion and, in the summer of 2017, I flew to Geneva for a 6-week internship at the European Organization for Nuclear Research (CERN), home to the world's largest particle accelerator, the Large Hadron Collider (LHC).
The Large Hadron Collider
This circular accelerator propels beams of particles to nearly the speed of light and collides them with each other at fixed points. These interaction points form cradles for the production of new, unstable particles like the Higgs boson, which was discovered with the LHC in 2012. One of these interaction points houses the Compact Muon Solenoid (CMS) experiment, which serves as a camera to record newly produced particles and their decay products and was home to the previous five years of my professional life.
Hunting for New Particles
After my return from Geneva, I stayed in touch with the UGent CMS group, where I completed my master's thesis. Eventually, in 2018, I started my PhD there. My main focus was on what we called âsearches for new physicsâ.
In the field of particle physics, we have a model called the Standard Model that describes all fundamental particles and how they interact. This model has been very successful in making accurate predictions confirmed by experiments. However, there are still some unanswered questions that the Standard Model can't explain.
One of the most popular questions revolves around dark matter, for example. The current Standard Model doesn't have any particles that can explain dark matter. Therefore, we know that the model is incomplete. To fill these gaps in our knowledge, scientists have suggested different additions to the Standard Model.
In particle collider experiments like CMS, we study collision data to search for new particles predicted by extensions to the Standard Model. We closely inspect known processes to detect any hints of unknown particles.
So what have I done?
During my time at CERN, I had the opportunity to work on two different aspects. Firstly, I was involved in the search for new physics, which was a significant part of my PhD thesis. This involved conducting thorough analysis projects from start to finish. I developed the analysis strategy, simulated signals, designed preprocessing techniques, estimated background and uncertainties, performed statistical analyses, and interpreted the results.
Additionally, I contributed to the maintenance of the experiment framework itself. To understand the context of this work, it is important to know that the LHC is designed to produce collisions every 25ns. If every collision event were saved, with about 1 MB of data per event, it would create around 40TB of data every second. Since we can only handle about 1,000 events per second, this is way too much data to manage on the fly.
Luckily many collision events at the LHC aren't important for CMS physics goals. Events useful for new physics analyses take place less than ten times per second. By filtering these crucial processes during data collection, we can decrease bandwidth to a manageable level. This is known as triggering. This filtering must happen very fastâtypically within a few hundred millisecondsâto make an immediate decision. During my PhD, I became an expert on a particular trigger type and took charge of its upkeep and improvements leading a team to accomplish this task.
Where does AI come in?
Identification of particles
When particles are created in the center of the detector, they undergo stochastic decay. Each type of particle has a distinct signature when it decays. You can imagine the CMS detector as a massive camera capturing the activity of these particles. We can then treat particle identification as a vision problem.
Let's consider an example: one of the jets shown in the image below. These jets originate from unstable particles called quarks, which subsequently cascade into a spray of other particles. Although the image below shows a 2D slice, in reality, the detector is cylindrical in shape, and we work within a 3D space with the center of our coordinate system located at the center of the detector.
To simplify analysis, we can preprocess data to determine an approximate position for each jet. By projecting this region onto a grid and utilizing it as input for a Convolutional Neural Network (CNN), we can effectively identify and analyze these particle jets.
Deep learning in triggers
In the project I led, our aim was to detect tau leptons during the triggering process. These are unique objects that decay into narrower jets with fewer particles than quarks.
Working in a low-timing environment presented its challenges, so we implemented multiple filters that use basic information to quickly eliminate obvious cases. This had the additional advantage of giving us an idea where to look for these particles in our detector. By taking this region and dividing it into a grid, we effectively translate our detector to a pixel space.
The image obtained in this pixel space is not an ordinary one. Instead of traditional color channels, it uses a channel to show the activity of each type of particle we observe. Nonetheless, we obtain an image on which we could train a CNN for further identification, a method that greatly improved the performance of the triggers.
AI in the Hunt for New Physics
When searching for undiscovered particles, it is crucial to distinguish between signal and background. Traditionally, this was done by creating mathematical quantities that represented physical properties of the target particle, like its mass. These variables were then used to construct cut-and-count strategies. However, in recent years, machine learning techniques have gained popularity in physics analyses.
Instead of relying on cut-and-count strategies, machine learning models can be trained to use both the same variables and low-level variables as inputs. This approach has been successfully applied in two analyses I worked on.
In one analysis, we utilized parametric neural networks. Since the undiscovered particles could have varying masses, we would have to train a different network for every possible mass value. This would be cumbersome and even unfeasible if we want to perform detailed mass analysis. Instead, we trained a single neural network where the mass information was treated as a parameter during training. This allowed us to effectively interpolate mass scenarios that were not previously seen.
The second analysis employed simpler machine learning techniques such as boosted decision trees to distinguish between signal and background. As these methods provided efficient and reliable discrimination, there was no need for complex neural networks in this study. This also made it simpler to understand the physics behind the model.
Life after my PhD
As my public PhD defense approached, I couldn't help but reflect on my journey. My time in particle physics at CERN has been nothing short of incredible. I've had the privilege to learn from brilliant minds and make lasting scientific contributions at the forefront of science.
However, as I delved deeper into my research, I discovered a new passion: AI. The endless potential and possibilities in this field drew me in, and I knew it was time to get involved. I decided I wanted to take my five years of experience in scientific data science, project management, and efficient and creative problem-solving skills to the industry and make positive and more short-term societal impacts.
What I was looking for
During my job hunt, I had a clear vision of what I was looking for in my new workplace. First and foremost, I wanted to be challenged by the projects I worked on, ideally at the same level as my previous experience at CERN. Additionally, it was important for me to find an environment that encouraged continuous learning and personal growth.
Considering that I was transitioning into a different field, finding a company that appreciated and valued my physics background as well as my outside-of-the-box thinking and problem-solving skills was crucial. Moreover, given my wide range of interests, I sought an opportunity where I could work on diverse problems rather than being limited to just one project.
Finally, I wanted to work in a place with a positive and ambitious atmosphere where teamwork matters. A place where everyone feels united but also has the freedom to manage their projects independently without strict guidelines. A place driven by the goal of making a positive and sustainable impact on the world.
A new chapter at Datameister
Datameister not only met but exceeded all these expectations. It is probably no surprise that when I heard about Datameister from a friend, I was immediately sold. After working here for about a month, my feelings haven't changed.
So that's itâthat's me and my journey in a nutshell. I'm excited about the future and can't wait to share more with you in upcoming blogs! đ