USCMS Researcher: Oz Amram

Postdoc dates: Sep 2022 - Sep 2024

Home Institution: Fermilab

Project: AI Denoising to Accelerate Detector Simulation

Develop fast and accurate simulation of highly granular calorimeters using machine learning models. Use a 'denoising' approach to upscale existing physics-based fast simulation to achieve better quality at similar speeds.

More information: My project proposal

Mentors:

• Kevin Pedro (Fermilab)

Presentations

• 8 Nov 2022 - "Machine Learning Models for Calorimeter Simulation", Oz Amram, HL-LHC R&D Initiative Meeting
• 11 May 2023 - "Fast and Accurate Calorimeter Simulation with Diffusion Models", Oz Amram, Computing in High Energy Physics (CHEP) 2023
• 30 May 2023 - "Denoising Diffusion Models for High Fidelity Calorimeter Simulation", Oz Amram, CaloChallenge Workshop

---

Current Status

2024 Q2

• Progress
  • Completed full data pipeline for training on CMS HGCal simulation.
  • Trained first models to reproduce single photon HGCalEE showers in a limited energy range at fixed rapidity and angle. Captured basic structure of the shower but modeling of some features can be improved
  • Developed first version of the model which uses existing fastsim as an input to reduce number of diffusion steps needed for high quality generation by using conditional distillation technique
• Next steps
  • Generate a larger set of CMS HGCal simulations covering a larger phase space. Use these simulations to train the model at a larger scale.
  • Improve the modeling of some HGCal shower features (energy per layer, radial spread of the shower, occupancy)
  • Develop testing suite to quantitatively validate performance of trained model
  • Further optimize usage of existing fast sim input to improve speedup gains

2024 Q1

• Progress
  • Began development of pipeline for training on CMS HGCal simulation, including handling complicated geometry.
  • Began development of hybrid model which incorporates existing fast simulation to reduce time needed for generation of high quality showers
• Next steps
  • Complete HGCal simulation pipeline, train first models.
  • Complete developement of hybrid model. Test speed up gains from its usage

2023 and prior

• Progress
  • Have successfully developed denoising diffusion model for calorimeter simulation that achieves state of the art results on public datasets.
  • Paper published in Phys.Rev.D 108 (2023)
• Next steps
  • Next steps are to explore different methods of improving the generation speed, including diffusing from existing fast simulation rather than pure noise.
  • Following this, we will apply the model to simulate the current CMS calorimeter and eventually the HGCal design.

Contact me:

• Send me an email