The Large Hadron Collider (LHC) is taking data at an unprecedented energy of 13.6 TeV, allowing to explore a new energy frontier and rare processes. In the meantime its high luminosity phase (HL-LHC) is actively being prepared. One of the main goals of both the on-going run and the HL-LHC phase is to scrutinize potential deviations from the predictions of the Standard Model of particle physics. For that purpose, it is particularly relevant to study the cornerstone of the model, the Higgs boson, as well as the heaviest known elementary particle, the top quark.
The PhD topic is aiming to develop machine learning techniques for two different challenging projects within the ATLAS experiment at the LHC: visual inspection of new ATLAS tracker modules, and reconstruction of final state objects to study rare processes with multiple leptons in the final state.

In order to meet the performance required by HL-LHC, the current ATLAS inner detector will be replaced by a new all-silicon detector, partially composed of pixel modules. The pixel module production includes multiple assembly steps each one being followed by quality control. The first part of the PhD project is aiming to automate visual inspection of the module metallic contacts (pads) before connecting them to the electronics. Initial work showed that visible defects presented on the pad surface are detectable by machine learning algorithms.

The PhD work will explore several directions:
- Full investigation of supervised approaches. The production start-up period will allow to collect the dataset necessary to train supervised segmentation models targeting not only anomaly detection, but also their classification.
- Investigation of unsupervised anomaly detection based on generative models. The DRAEM model [1] for instance learns a joint representation of an anomalous image and its anomaly-free reconstruction, while simultaneously learning to distinguish them. The reconstructive sub-network is formulated as an encoder-decoder architecture and can be replaced with more advanced types of generative models such as GANs.
- Investigation of various models treating an anomaly-free pad surface as a background noise. Signal processing denoising algorithms can then be used to increase the information content while e.g denoising autoencoders can learn features that are invariant to noise.

The second part of the PhD will focus on rare processes involving top quarks with bosons (ttW, ttZ, ttH, tH, 4tops). These processes have just been observed for the first time using the dataset recorded at 13 TeV (except for tH) in final states with several leptons. However these analyses have always been carried out separately, which strongly limits our overall physics understanding, the correlations between the various measurements being large and unknown. The new approach proposed here consists of designing a global analysis to tackle these limitations. This approach requires cutting edge techniques to reconstruct precisely the final state. Indeed several degrees of freedom originate from the undetected neutrinos and combinatorics is made complex by the multiple decay products. An approach based on GNN was already tried. Although it provides a certain level of accuracy, some other approaches have to be explored to improve the reconstruction performance. There are in particular two models of interest developed within the LHC community:
- The topograph model [2], also based on GNN, makes use of additional information about the event decay chain.
- The Spanet model [3], a symmetry-preserving attention network exploits natural invariances to efficiently find particle assignments without evaluating all permutations.
An assessment of the relevant uncertainties is also an important aspect when using these tools.

This dual project provides an opportunity to be involved in the full chain of the AI application development, being also part of the world-wide ATLAS collaboration.