Focus on:  Hide Contributions
Back to Conference View
Login

LPCC Fast Detector Simulation Workshop

→ Europe/Zurich
Alberto Ribon (CERN), Anna Zaborowska (CERN), Witold Pokorski (CERN)
Description

The goal of the workshop is to discuss the developments of the fast detector simulation and to identify common interests with room for collaboration. The workshop has been divided into 6 sessions to facilitate the discussions on the topic in focus. Please find the list of useful questions around the topic in the description of each session in the timetable.


The Workshop will be held in remote mode only.

Participants
105 View full list
LPCC Fast Detector Simulation Workshop
Zoom Meeting ID
68574150571
Host
Anna Zaborowska
Useful links
Join via phone
Zoom URL
  • Monday 22 November
    • 1
      Introduction
      Speakers: Alberto Ribon (CERN), Anna Zaborowska (CERN), Witold Pokorski (CERN)
    • Fast Simulation applicability and accuracy

      How is fast simulation used: applied to which detectors? Are there specific โ€œsensitiveโ€ parts of the detector where it is difficult to get fast simulation? Is it used widely in production? Which physics analyses can use fast simulation, which cannot? What are the demands on the accuracy of the (fast) simulation? Do you/can you tune fast simulation to collision data to get better results than from the full simulation?

      How does it look now, but also how it changes for Run3 & HL-LHC?

      • 2
        CMS

        Session: Applicability and accuracy

        Speaker: Samuel Louis Bein (Hamburg University (DE))
        LPCCFastSimApplicability.pdf
      • 3
        ATLAS

        Session: Applicability and accuracy

        Speakers: Joshua Falco Beirer (Georg August Universitaet Goettingen (DE)), Joshua Falco Beirer (CERN, Georg-August-Universitaet Goettingen (DE)), Joshua Falco Beirer (Georg August Universitaet Goettingen (DE))
        LPCC_2021_Beirer.pdf
      • 4
        ALICE

        Session: Applicability and accuracy

        Speaker: Roberto Preghenella (INFN, Bologna (IT))
        LPCC Fast Sim.pdf
      • 5
        LHCb

        Session: Applicability and accuracy

        Speaker: Constantine Chimpoesh
        [LPCC]-Generative_models_uncertainty_estimation.pdf
      • 6
        Discussion
    • Data preprocessing

      Which particle types and properties (energy, angle) are considered?
      How large is the full sim dataset used to extract parametrization/train the network? Is the dataset balanced (is the number of events for the different particle properties almost the same?)
      How the structure of the input data is defined (hits, cells, clusters, custom voxels,..)?
      Which data structure is used for the ML training (1D vector, images, graphs..)?
      Is input data scaled? How?
      How do you store the preprocessed data?
      How are the condition values for the ML training (energy of the particle, angle,..) encoded?

    • 16:55
      Break
    • Tuning

      How the extraction and tuning of parameters for the fast simulation are done?
      For ML, how training is performed? What hyperparameter tuning approach is adopted? What metric is used to compare the performance of the model during the tuning process?

    • 18:00
      Further discussions

      In case we need more time.

  • Tuesday 23 November
    • Validation

      How do you validate fast simulation?
      Is it on e.g. particle shower characteristics (longitudinal/transverse profiles, etc.), on high-level results of physics analyses, ...?
      For ML, what is the metric used to validate the model and what is the objective function you are optimizing during the training?

    • Fast sim model integration

      How is fast simulation integrated into the experiment's framework?
      How is the model stored (file format, size, ...)? Are there any specific files to store in addition to the model?
      For ML, which inference library is used?
      How much time does it take to generate one event?
      What is the memory footprint?
      Are there any optimization approaches considered to reduce the memory?
      Has the ML simulation been tested on GPU?

    • 16:50
      Break
    • Reusability and Generalisation

      Are there any valuable lessons to be learnt by others?
      Tools that can be reused?
      Maybe some ideas have failed for your detectors (or parts of it) and it can be useful to others?
      What are your plans, maybe there is room for collaboration?

    • 18:05
      Further discussions

      In case we need more time.

Powered by Indico v3.3.7-pre