As an experimental particle physicist, my goal is to study the fundamental building blocks of Nature. The standard model (SM) of particle physics, the theory that describes the forces that regulate the interaction between the elementary particles, has provided a very successful description of the Universe, validated by decades of experiments. As a PhD student I contributed to one of these validations. As a member of the Collider Detector at Fermilab (CDF) experiment, that operated at the Tevatron collider at Fermi National Accelerator Laboratory in Chicago, I had the opportunity to co-lead the effort that discovered one of the predicted but at the time still unobserved production mechanisms of the top quark, the heaviest of all elementary particles. Despite the success of the SM, many questions at the forefront of particle physics remain unanswered, one of the most important being the nature of dark matter (DM). Experiments at the Large Hadron Collider (LHC) at CERN in Geneva, are primarily positioned to address some of these questions.

Although cosmological observations demonstrate that around 85% of the mass of the Universe is comprised of DM, details of its composition remain elusive. If DM particles interact with SM particles, DM can be produced by colliding protons at high energy. While the LHC is at the energy frontier of colliding beams, its experiments were not designed to directly detect DM: once produced, DM particles would not interact with the detectors, passing through them without leaving any trace. However, if they are produced in association with other detectable particles, they can manifest themselves as an imbalance in the total momentum. The estimate of the momentum that has been carried away by non detectable, “invisible” particles, but that is expected due to the law of conservation of energy is called missing transverse momentum (MET). At the LHC, an entire program has been built on searches for DM in collisions that produced large MET. As a member of the Compact Muon Solenoid (CMS) experiment, I contributed to this program publishing two searches for DM produced in association with a top quark and in association with a SM Higgs boson. 

I was also selected as the CMS delegate to the LHC DM Working Group, a cross-experiment effort that defines a set of shared guidelines and recommendations for the LHC physicists on interpretation, presentation, and characterization of DM results.

The LHC DM program on searches with large MET is currently limited by the performance of the MET reconstruction. To improve the MET that is used for the offline analysis of the data collected by CMS, I am leading the development of a graph neural network (GNN) for MET reconstruction. My work took inspiration from existing machine learning (ML) MET reconstruction methods and showed a significant and promising improvement. From 2020 to 2022 I also led the MET group of the CMS experiment, where I promoted the adoption of ML techniques while simultaneously maintaining and improving the more classical MET reconstruction approaches. Although a crucial aspect, the resolution of the MET used for offline analysis is not the only bottleneck for MET-based DM searches. At the LHC, millions of collisions are delivered every second, but only a fraction of those can be stored by the experiments, due to the limited bandwidth, processing, and storage resources. In collider experiments, a mechanism called the "trigger" classifies and selects events in real-time. It is composed of a combination of algorithms, implemented in both hardware and software, that save only the events with characteristics of interest as they occur. The hardware component has only a few microseconds to decide whether to record events. Due to these extremely low latencies, the algorithms used at this stage are deployed on custom electronics, like field-programmable gate arrays (FPGAs). My plan is to develop tools for the deployment on FPGAs of GNNs for MET reconstruction similar to the one developed for offline analysis. Such tools will be rather general and could be applied to the deployment of GNNs developed for other scopes. Furthermore, while microsecond latency requirements are unique to CMS, the ability to perform real-time analysis with high efficiency could be relevant to other applications in physics and beyond. Real-world applications like on-device image processing performed by intelligent cameras or the navigation of autonomous vehicles may benefit from the same techniques.