Thesis

At CERN, the LHCb experiment sees protons collide tens of millions of times per second and must decide, in real time, which collisions to keep. Today this is done by thousands of small programs, each hand-written for one physics signal, an approach that no longer scales. Meanwhile, AI has been revolutionised by “foundation models”, the kind of large models behind tools like ChatGPT: train one big model once on huge amounts of data, then reuse it for many different tasks. This PhD asks whether we can build such a model for LHCb: a single AI that learns to “read” a whole collision and can then be reused across many physics studies. You would build the first version of this model, train it directly on real CERN data (not just simulations, which is what makes it novel), and show that it transfers from one task to another while needing far less hand-labelling. Along the way you will learn modern AI hands-on (PyTorch, large-scale GPU training) and real data analysis from a flagship experiment. No prior AI expertise is required, curiosity, solid programming, and an interest in physics or machine learning are what matter. The work is hosted at CEA/Irfu, alongside the HEPfoundry team building foundation models for physics, with partners at CERN/LHCb.