I'm a postdoctoral researcher specializing in experimental High-Energy Physics (HEP), focusing on the study of fundamental particles of matter and their interactions. With over a decade of experience, I've worked extensively on the CMS experiment at the Large Hadron Collider (LHC) in Geneva, Switzerland.
I'm proficient in software development & advanced computing tools, such as High-Performance Computing (HPC) & Machine Learning (ML), for investigating the big data sets recorded at the LHC.
I enjoy tackling complex problems and solving technical challenges while continuously expanding my skill set.
The LHC experiments aim to measure the properties of fundamental particles, building blocks of matter, that existed only in the early universe, thereby expanding our understanding of particle physics.
My research has been focused on the following topics:
More details can be found in my publication list.
The LHC experiments are Big Data generators, \(\mathcal{O}\)(Exabytes), requiring sophisticated algorithms and statistical methods for quantitative analysis in particle physics. This process extensively employs machine learning techniques as well. Drawing on my experience, I led the CMS Machine Learning Innovation Group (2020–2022) and since 2024 the Inter-Experimental LHC ML Group at CERN.
Throughout my career I acquired extensive knowledge in various areas:
I use C++ & Python daily and am experienced in working with and contributing to large codebases. I am proficient at learning and adapting to new programming languages and concepts independently.
A detailed list of my software expertise can be found in my CV. Some of my smaller developments can be found on github.
Working on particle physics detectors involves a blend of advanced technology, innovative engineering, and a deep understanding of physics, which I apply to the development of the high granularity calorimeter (HGCAL) upgrade of the CMS detector.
I have a keen interest and posses expertise in the following areas:
Neural network to search for exotic long-lived particles
CMS experiment at the CERN LHC
I am leading efforts in searching for 3 or 4 top quark events via heavy beyond the standard model (BSM) resonances, while simultaneously providing guidance to two PhD students engaged in related analyses and machine learning projects. Additionally, I am actively involved in characterizing HGCAL SiPM-on-tile modules intended for the CMS detector upgrade at the CERN LHC.
I conducted extensive research on long-lived heavy neutral leptons decaying to leptons and displaced jets, while overseeing the progress of two PhD students throughout the study. Moreover, I introduced a novel methodology to measure the top quark width, utilizing a neural network to infer the b quark charge for distinguishing between on- and off-shell production.
I searched for long-lived particles using a novel displaced jet tagger based on state- of-the-art neural network techniques. Furthermore, I exploited neural networks for improving the sensitivity in measurements of B hadron decays and in the FPGA-based event trigger system.
I steered the development of the VISPA@web application, a graphical development environment for high energy and astroparticle physics analyses.
In this role, I serve as representative of the CMS collaboration. Together with other representatives we organize monthly meetings to facilitate and foster an exchange of ML ideas between the LHC experiments (ATLAS, CMS, LHCb, ALICE), theory, and accelerator groups. At our annual workshop we regularly connect with experts from industry (eg. from DeepMind, Meta, etc) as well. Further information can be found on our website.
In this role, I regularly organized meetings to review the methods and progress of approximately 30 analysts engaged in single top quark research, ensuring effective communication and collaboration to facilitate publications.
The group is responsible for identifying and integrating new and advanced machine learning (ML) developments into all areas of the CMS experiment, including data analyses, detector simulations & calibration, data quality monitoring, and the event trigger system. I coordinated topical hackathons (15-25 participants per hackathon) and managed regular seminars (about 20 participants) to stay updated on the latest ML advancements in the field.
In this role, I was responsible for overseeing the statistical inference and the treatment of systematic uncertainties in the precise measurements of parameters of the Standard Model of particle physics and of parton distribution functions.
In this role, I assumed responsibility for the dedicated track reconstruction used in the CMS FastSimulation package, managing and reviewing the work of a team of five developers.
Recipient of the FRIA grant (fully-funded doctoral scholarship). Awarded by Fonds National de la Recherche Scientifique (F.R.S.-FNRS), Belgium.
PhD thesis: "Differential single-top-quark cross sections in t channel at 8 and
13 TeV with the CMS experiment".
Supervisor: Prof. Dr. Andrea Giammanco.
Final mark: excellent (100%).
Master thesis: "Measurement of top-quark spin asymmetries in t-channel
single-top-quark production at 7 TeV with the CMS experiment".
Supervisor: Prof. Dr. Martin.
Erdmann.
Final mark: very good (91%).
Bachelor thesis: "Development of the VISPA@web program and
measurement of the dijet mass with CMS".
Supervisor: Prof. Dr. Martin Erdmann.
DESY summer student programme, Hamburg, Germany.
CMS Data Analysis School (DAS), CERN, Switzerland.
Terascale school for undergraduate students, Hamburg, Germany.
CMS Data Analysis School (DAS), Hamburg, Germany.
Imperial College London, UK.
3rd CMS FastSim workshop, Fermilab, USA.
RWTH Aachen University, Germany.
Conference on Computing in High Energy and Nuclear Physics (CHEP'19)
A web-based bokeh application to inspect data recorded during test beams with CMS HGCAL SiPM-on-tile modules. The application is written in Python and uses uproot to process CERN ROOT files. Since the visualization is within a browser the data on the server-side does not need to be on the same system as the client.
A set of C++ plugins to asynchronous read and process CERN ROOT files directly within the multithreaded TensorFlow (v1) data pipeline for neural network training.
Unfortunately, the input pipeline has been significantly restructured in TensorFlow v2 such that these plugins won't work with newer versions.
This repository shows the training of a deep neural network multiclass classifier. The b jet charge is inferred from particle jet constituents including charged particle tracks, neutral particles in the calorimeters, muons, and electrons, as well as secondary vertices formed by at least two tracks. The resulting score is charge-symmetric by construction, ie. flipping all charges of the tracks will also flip the predicted class.
A simple CMSSW plugin to artificially generate long-lived particles and their isotropic decay. The particle types, lifetime, and decay channels can be easily adapted to emulate various signatures.
This plugin is of great use for generating large samples of artifical particles to improve the generalization of neural network classifiers.
This project targets the reconstruction of the primary interaction vertex using a multistage neural network. A novel aspect is the use of a custom TensorFlow plugin to generate histograms from tracks with a learnable weight function.
Since the final classifier is ported into FPGAs a high focus on its computational footprint (eg. BRAM, DSP, etc.) and latency is required.
A C++ based CMSSW plugin to convert collections in the CMSSW-internal EDM format into the PXLIO format used by the VISPA project.
template<class T> void ComputeTmpl( OpKernelContext* context, unsigned int input_index, const std::vector<unsigned int>& diced_rates, unsigned int sum_rates ) const { const Tensor& input_tensor = context->input(input_index); auto input_data = input_tensor.flat<T>(); int batch_size = input_tensor.dim_size(0); int batch_length = input_tensor.NumElements()/batch_size; Tensor* output_tensor = nullptr; TensorShape output_shape = input_tensor.shape(); output_shape.set_dim(0,sum_rates); OP_REQUIRES_OK(context, context->allocate_output(input_index-1, output_shape,&output_tensor)); auto output_data = output_tensor->flat<T>(); unsigned int output_batch_index = 0; for (unsigned int ibatch = 0; ibatch<batch_size; ++ibatch) { for (unsigned int icopy = 0; icopy<diced_rates[ibatch];++icopy) { for (unsigned int idata = 0; idata<batch_length; ++idata) { output_data(output_batch_index*batch_length+idata) = input_data(ibatch*batch_length+idata); } output_batch_index+=1; } } }
Custom TensorFlow kernel for data preprocessing
Die shot Intel i9-13900K by Fritzchens Fritz
CMS Collaboration, JHEP 03 (2024) 105 Search for long-lived heavy neutral leptons with lepton flavour conserving or violating decays to a jet and a charged lepton
The LHC collision data has been searched for signs of exotic long-lived particles, so-called heavy neutral leptons. Their existence could answer various observed phenomena: Why are the neutrino masses so tiny? What is the nature of dark matter? Why is there much more matter than antimatter in the universe?
Unfortunately, no evidence for such particles have been found, yet the search significantly narrowed their possible parameters (eg. the particle's lifetime \(c\tau_{0}\) as shown in the plot) helping future analyses to focus their efforts.
CMS Collaboration, Mach. Learn.: Sci. Technol. 1 (2020) 035012 A deep neural network to search for new long-lived particles decaying to jets
An advanced deep neural network (DNN) has been developed, designed to tag jets, sprays of collimated particles, that originate from 0.1mm and up to 1m away from the proton-proton collision point in the CMS detector's center. This novel classifier significantly enhances our sensitivity in searching for exotic long-lived particles (LLPs), which can travel macroscopic distances before decaying, as illustrated in the accompanying plot.
A key innovation is the use of domain adaptation. This technique ensures that the performance of the classifier, which is initially trained on simulated data, matches its performance on real data. This on-the-fly calibration ensures that the DNN predictions remain accurate and reliable in practical applications.
CMS Collaboration, Eur. Phys. J. C 80 (2020) 370 Measurement of differential cross sections and charge ratios for \(t\)-channel single top quark production in proton-proton collisions at \(\sqrt{s}\) =13 TeV
The elusive production of single top quarks was first observed in 2009, at the Tevatron collider. Now, for the first time, the LHC enables comprehensive studies of single top quark production. This paper includes also a first measurement of the production ratio of top quark (\(t\)) and its antiparticle (\(\bar{t}\)), providing insights into the quark composition of the proton.
The measured ratio of about \(\frac{2}{3}\), as shown in the figure, confirms that the proton consists of two up quarks and one down quark. This is because interacting up quarks can lead to the production of top quarks, whereas interacting down quarks yield their antiparticles.
CMS Collaboration, JHEP 04 (2016) 073 Measurement of top quark polarisation in \(t\)-channel single top quark production
A distinctive feature of single top quarks is their near 100% polarization. Simply put, this means their spin is almost always aligned with their direction of flight. The accompanying figure illustrates an approximation of the cosine of the polarization angle. The degree of polarization can be derived from the slope of the data points.
The polarization provides valuable insights into the properties and interactions of single top quarks. Such results can also indirectly exclude certain new theories, that postulate new interactions and/or particles. If these would cause a depolarization of top quarks that contradicts the observed data, they can be ruled out.
CMS Collaboration, CMS-PAS-TOP-16-004 Measurement of the differential cross section for \(t\)-channel single-top-quark production at \(\sqrt{s}\) =13 TeV
Previously, the LHC operated at around half its maximum energy, specifically at 7 and 8 TeV. In 2015, it achieved a significant milestone by increasing its energy to 13 TeV for the first time. This paper presents the initial measurement of single top quark production using the very first data collected at this unprecedented energy level.
The figure displays the output score of a boosted decision tree (BDT), which effectively separates single top quark events (shown in red and labeled "\(t\)-channel") from other processes.
European Physical Society Conference on High Energy Physics (EPS-HEP), Hamburg, Germany. Proceedings: PoS (EPS-HEP2023) 452
11th International Workshop on the CKM Unitarity Triangle (CKM), (virtual). Proceedings: PoS (CKM2021) 024
9th Annual Large Hadron Collider Physics Conference (LHCP), (virtual). Proceedings: PoS (LHCP2021) 001
24th International Conference on Computing in High Energy and Nuclear Physics (CHEP), Adelaide, Australia. Proceedings: EPJ Web Conf. 245 (2020) 06013
39th International Conference on High Energy Physics (ICHEP), Seoul, Korea. Proceedings: PoS (ICHEP2018) 025
10th International Workshop on Top Quark Physics (TOP), Braga, Portugal. Proceedings: arXiv:1711.11445 [hep-ex]
QCD@LHC, Zurich, Switzerland.
4th Annual Large Hadron Collider Physics Conference (LHCP), Lund, Sweden. Proceedings: PoS (LHCP2016) 168
9th International Workshop on Top Quark Physics (TOP), Olomouc, Czech Republic. Proceedings: arXiv:1611.04852 [hep-ex]
8th International Workshop on Top Quark Physics (TOP), Ischia, Italy. Proceedings: PoS (TOP2015) 023
19th International Conference on Computing in High Energy and Nuclear Physics (CHEP), New York, USA. Proceedings: J.Phys.Conf.Ser. 396 (2012) 052015
CMS Upgrade Days, CERN, Switzerland.
Spring Conference of the German Physical Society (DPG), (virtual).
PyHEP Workshop, (virtual).
Invited talk at VBS Polarization Workshop, Laboratoire Leprince-Ringuet (LLR), France.
5th CMS Single-Top Workshop, Oviedo, Spain.
4th CMS Single-Top Workshop, Karlsruhe, Germany.
3th CMS Single-Top Workshop, Strasbourg, France.
Seminar at University of Hamburg, Germany.
1st CMS Single-Top Workshop, Naples, Italy.
Worldwide LHC Computing Grid for distributed data analyses