Machine Learning based fast simulation tools

R&D task number: G4RD7

Monte Carlo simulation is vital for the design, construction and operation of high-energy physics (HEP) experiments. The computing resources needed to generate the required amount of simulated data are increasing with the energy and the luminosity of particle accelerators and are starting to exceed the available computing budget. Faster simulation is therefore essential to maintain the accuracy of physics analyses. Speed-up of the simulation may be achieved by reducing the time spent in detailed simulation, by applying a parametrisation of the physics processes, or by directly reproducing the signal created in the detectors. With Machine Learning (ML)-based models, new possibilities have emerged where trained generative models replace the most compute-intensive stages of HEP event simulation. They are of special interest for the simulation of calorimeters, where most of the simulation time is spent for HEP experiments.

Technical objectives

• Development generalizable and fast adaptive generative models based on meta learning approaches for particle shower simulation.
• Integration of machine learning aided simulation into the Geant4 simulation
• Development optimized data pipelines from generating the full simulation events to the model training, valiation and optimization to the inference in C++
• Review and evaluation of inference libraries
• Review, evaluation and integration of memory footprint optimization strategies of the fast simulation model (training and inference)

For more information please contact Anna Zaborowska (anna.zaborowska@cern.ch) and Dalila Salamani (dalila.salamani@cern.ch).

Website: https://g4fastsim.web.cern.ch/