The Large Hadron Collider (LHC) at CERN is the world’s most powerful particle accelerator, colliding protons at very high energies. One of its main experiments, the ATLAS detector, records the particles produced in these collisions to test the Standard Model and search for new physics.
ATLAS analyses depend strongly on simulated data. Simulation helps us understand how the detector responds to particles, estimate backgrounds to signals of interest, and provide training data for machine learning algorithms. To produce this simulated data, theoretical models of particle interactions are combined with Monte Carlo event generators, which create lists of particles emerging from the collisions. These particles are then passed through a “digital twin” of the ATLAS detector, which models how they interact with the detector’s materials and sensors. The resulting output is structured like real data, so the same software can be used for both.
This detector simulation, typically done with the Geant4 toolkit, is very accurate but extremely slow, especially for the calorimeters that measure the energies of photons, electrons, and hadrons. For complex events, a single fully simulated event can take minutes of CPU time, limiting the number of events that can be produced. With the High-Luminosity LHC (HL-LHC) upgrade starting around 2030, collision rates will increase by about a factor of seven, making fast simulation even more critical.
Machine learning–based generative models offer a promising solution. Techniques such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and Normalising Flows can learn to reproduce complex particle shower patterns from training data and then generate realistic synthetic events much more quickly than traditional simulation. This project will develop and test such models for photon showers in the ATLAS calorimeters, aiming to achieve large speed gains while maintaining sufficient accuracy for physics analyses. Careful validation against standard Geant4 simulation and real collision data will be essential to ensure that these fast simulations can be reliably used in future ATLAS studies at the HL-LHC.
Below are some milestones of the proposed project:
M1: Perform a focused literature review of state-of-the-art generative AI methods (e.g. GANs, VAEs, Normalising Flows) and identify the most promising techniques for simulating realistic photon showers in calorimeters.
M2: Explore and characterise an existing pre-processed dataset of photons interacting in the ATLAS electromagnetic calorimeter, including its features, preprocessing steps, and relevant physics observables.
M3: Design and implement a robust data pipeline that converts this dataset into a suitable format for training generative models, including data loading, normalisation, and train/validation/test splits.
M4: Train, validate, and tune the selected generative models on the dataset, systematically assessing their ability to reproduce key shower properties and correlations.
M5: Benchmark the new models against existing ATLAS GAN-based calorimeter simulations, using common physics and technical performance metrics.
M6: Extend the training to larger internal ATLAS datasets prepared in advance, ensuring scalability of the models to more realistic data volumes and conditions.
M7: Evaluate the final models in detail, comparing physics fidelity (realism of the generated showers) and computing performance (generation speed and resource usage) against both traditional Geant4 simulation and current AI-based approaches.
M8: Summarise the overall findings and draw conclusions about the suitability and potential impact of the chosen generative techniques for fast calorimeter simulation in ATLAS and HL-LHC workflows.
Courses relevant from the bachelor's degree:
Recommended courses for theoretical curriculum in the master's programme:
(If not taken during bachelor)