Parts | Days | Selection | Search | Updates | Downloads | Help
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 4: AKPIK III: Simulation & Application
AKPIK 4.1: Talk
Thursday, March 18, 2021, 16:00–16:15, AKPIKa
Deep Learning for Accelerating High Energy Physics Simulations — •Florian Rehm1,2, Sofia Vallecorsa1, Kerstin Borras2,3, and Dirk Krücker3 — 1CERN (Switzerland) — 2RWTH Aachen University (Germany) — 3DESY (Germany)
In particle physics the simulation of particles transport through detectors require an enormous amount of computational resources. This motivated the investigation of different, faster approaches, to replace the standard Monte Carlo. We use Generative Adversarial Networks to simulate electromagnetic calorimeter responses and decrease the simulation time by orders of magnitudes while keeping the necessary level of accuracy. The standard approach for generating 3D images uses 3D convolutional layers, however, 3D convolutional networks are demanding in terms of computational resources and memory. We present a novel architecture using 2D convolutional layers for representing the 3D images which reaches a higher accuracy and reduces the computational time by a factor of 3. We further reduce the inference time by quantizing the neural network parameters to a lower precision using the novel Intel low precision optimization tool (iLoT). Performance benchmarks on Intel Xeon processors yield a 1.73x speed-up. Particle simulations follow the rules of quantum field theory. Therefore, it is reasonable to explore the potential of quantum computers for these simulations. However, today's quantum computers are by far not capable to solve such complicated tasks. Hence, the further planned initial investigations employ simplified models to study the performance of quantum computers.