Soutenance de thèse de doctorat de Muhammad Usman

Machine Learning Driven Search for New Physics and Luminosity Measurement in pp Collisions at sqrt(s) = 13 TeV Using the ATLAS-TPX Network

The ATLAS experiment at the Large Hadron Collider provides an unparalleled setting for probing the predictions of the Standard Model and searching for signs of new physics beyond it. This thesis brings together two complementary lines of research that strengthen both the precision and discovery capabilities of the experiment. The first advances the measurement of luminosity through the ATLAS-TPX detector network, while the second introduces an innovative machine-learning approach, BumpNet, for model-independent resonance searches in invariant-mass spectra. Together, these efforts contribute to the broader ATLAS program by strengthening the experiment’s luminosity measurement program and advancing data analysis methodologies essential for the High-Luminosity LHC era.

The ATLAS-TPX network comprises fifteen pixelated silicon detectors based on Timepix ASICs, installed in the ATLAS cavern to monitor both the radiation field composition and luminosity throughout Run 2. Each two-layer detector is equipped with neutron converters and operates in two modes: Timeover-Threshold (ToT) and Time-of-Arrival (ToA), enabling three distinct algorithms for luminosity determination: Cluster Counting, Hit Counting, and Total Deposited Energy. After cross-calibration with the LUCID luminometer, the system delivers approximately 150 relative luminosity measurements per run. Comparisons with other ATLAS luminometers demonstrate good overall consistency, while small deviations motivated detailed investigations into systematic effects such as track overlapping corrections. Long-term stability analyses covering the full 2016–2018 Run 2 period confirm the reliability and durability of the detectors. This work directly supports the ATLAS Luminosity Working Group by improving redundancy, validating cross-calibrations, and reducing systematic uncertainties in luminosity, an essential input for all precision measurements and cross-section determinations.

The second part of this thesis presents BumpNet, a convolutional neural network (CNN) designed to automate and generalize the search for resonant structures (“bumps”) in invariant-mass histograms. Traditional bump hunts often rely on fixed background models or signal templates, limiting their applicability to specific hypotheses. BumpNet circumvents these constraints by predicting local loglikelihood significance values directly from smoothly falling spectra, enabling fast and model-agnostic searches across many final states without explicit background estimation. Trained on analytically generated and realistic datasets, the network reproduces the expected Higgs boson significance and agrees well with ATLAS dilepton resonance results, demonstrating its robustness and interpretability. Ongoing studies use ATLAS Run 2 Monte Carlo simulations to further optimize performance before deployment on real collision data. Within the ATLAS Machine Learning and New Physics Working Groups, this approach exemplifies how deep-learning-based anomaly detection can accelerate data analysis pipelines and extend the experiment’s sensitivity to unforeseen physical signatures.

Overall, the work in this thesis enhances the precision and discovery reach of the ATLAS experiment. Through the ATLAS-TPX network, it contributes to relative luminosity measurements; through BumpNet, it pioneers an adaptable, data-driven method for new-physics searches. Both efforts advance ATLAS’s strategic objective of maximizing discovery potential and readiness for the forthcoming High-Luminosity LHC program.