In high-energy physics experiments such as the Large Hadron Collider (LHC), some particles interact electromagnetically and/or hadronically with the material of the calorimeter, creating cascades of secondary particles, or showers. Describing the showering process relies on simulation methods that precisely define all particle interactions with matter. A detailed and accurate simulation is based on the Geant4 toolkit. The simulation of showers, with large amounts of particles created and tracked, is inherently slow. Alternatively, machine learning techniques such as generative models are used to speed up the generation of showers in a calorimeter, i.e., simulating the calorimeter response to certain particles.
Considering this, we are investigating a few different kinds of generative models such as VAE, VQ-VAE, and Diffusion based on transformer architecture. Diffusion models have proven to be significantly more accurate than others, which is what we need. However, these diffusion models come at the cost of slow inference. Therefore, this project aims to make the inference of diffusion models faster.
Furthermore, a byproduct of this project is that the student will get to work with diffusion transformer models which are currently at the forefront of AI research and learn to use them in the context of high-granularity shower data (a part of CaloChallenge).
A fast and accurate diffusion model for fast simulation of calorimetry showers.
(Please don’t email before completing the test. It’s not possible to reply to everyone.)