Generative Models for Application-Specific Fast Simulation of LHC Collision Events

Print Friendly, PDF & Email

In this video from PASC18, Maurizio Pierini from CERN presents: Generative Models for Application-Specific Fast Simulation of LHC Collision Events.

“We investigate the possibility of using generative models (e.g., GANs and variational autoencoders) as analysis-specific data augmentation tools to increase the size of the simulation data used by the LHC experiments. With the LHC entering its high-luminosity phase in 2025, the projected computing resources will not be able to sustain the demand for simulated events. Generative models are already investigated as the mean to speed up the centralized simulation process. Here we propose to investigate a different strategy: training deep networks to generate small-dimension ntuples of numbers (physics quantities such as reconstructed particle energy and direction), learning the distribution of these quantities from a sample of simulated data. In one step, one would then be able to generate the outcome of the full processing workflow (generation + simulation + reconstruction + selection).”

Maurizio Pierini is the coordinator of the Physics Performances and Dataset (PPD) area at CERN. “Our task is to provide CMS with the detector conditions (alignment and calibrations), assure the quality of our data (online and offline data quality monitoring and certification), and to validate the performances of the software for generation, simulation, reconstruction, and physics-object definition. It is our task to coordinate the Monte Carlo production and the dataset (re)processing, including the definition of primary datasets, and skims for physics and detector studies. Being responsible of the datasets, we also take care of data-flow related applications like hotline, event display, and scouting.”

Co-Author(s): Dominick Olivito, Bobak Hashemi, Nick Amin (UC San Diego)