Scientific simulation calculations are often limited by their high demand for computing capacity. Generative Deep Neural Networks offer an efficient way to replace complex models and enable fast and precise simulations for the CMS and ATLAS experiments at the Large Hadron Collider (CERN).
How can AI models help in modelling the energies and positions of charged and neutral particles at the Large Hadron Collider? Read in this week’s Helmholtz AI project showcase how researchers at the Deutsches Elektronen-Synchrotron (DESY) and Forschungszentrum Jülich (FZJ) are working on this challenge.
Could you introduce yourself, giving your affiliation, area of work, and of course, the project title?
Hello, my name is Dirk Krücker and I work at DESY in the field of particle physics and participate in the Compact Muon Solenoid (CMS) Experiment, a collaboration that operates the CMS detector at the Large Hadron Collider (LHC) at CERN/Geneva. In recent years, my focus has been on the application of Deep Learning in the search for New Physics. In the DeGeSim project at DESY, Kerstin Borras, Judith Katzy and Moritz Scham are involved. In addition, we cooperate with Jenia Jitsev from the Jülich Supercomputing Centre (JSC) at FZJ and with our associated partner Wojtek Fedorko at TRIUMF (Vancouver). I also participate in the Helmholtz AI project SynRap.
In simple words, what specifically is your project about? And, how and why do you think it is a high risk, high gain endeavour?
The DeGeSim project consists of two work packages on generative modelling with Generative Adversarial Networks (GANs), in particular GraphNNs. J. Katzy from the ATLAS experiment, also at the LHC, is interested in modelling the response of the ATLAS calorimeter to certain background processes, the so-called pile-up. The CMS-related part deals with the generation of artificial data for the planned High Granularity Calorimeter (HGCal). The HGCal consists of partially irregular hexagonal and partially grid-like sensors. This complex geometry does not map well to conventional CNNs. The system provides about 2 x 3.1 million channels with high dynamic range. Presently, no generative model has been developed for such complex calorimeters, and it is unclear whether known approaches scale. However, fast simulation techniques are essential for current and future collider experiments. All future data analyses at the LHC will benefit from a breakthrough in these techniques.
How important has the Helmholtz AI funding and platform been to carry out this project?
Essential, since it allows me to dedicate a larger part of my time together with two students on the project.
What else would you like to share?
The project brings together the Deep Learning expertise of FZJ/JSC with DESY's experience in particle physics. We are happy to have TRIUMF as an associated partner. Our colleagues in Vancouver are also exploring the modelling of electromagnetic calorimeter showers.