Open AccessBENCHMARK OF GENERATIVE ADVERSARIAL NETWORKS FOR FAST HEP CALORIMETER SIMULATIONS
Author(s) -
Florian Rehm,
S. Vallecorsa,
K. Borras,
D. Krücker
Publication year - 2021
Publication title -
9th international conference "distributed computing and grid technologies in science and education"
Language(s) - English
Resource type - Conference proceedings
DOI - 10.54546/mlit.2021.30.99.001
Subject(s) - detector , computer science , benchmark (surveying) , monte carlo method , large hadron collider , convolutional neural network , deep learning , calorimeter (particle physics) , computer engineering , artificial intelligence , algorithm , computational science , physics , particle physics , telecommunications , statistics , mathematics , geodesy , geography
Highly precise simulations of elementary particles interaction and processes are fundamental toaccurately reproduce and interpret the experimental results in High Energy Physics (HEP) detectorsand to correctly reconstruct the particle flows. Today, detector simulations typically rely on MonteCarlo-based methods which are extremely demanding in terms of computing resources. The need forsimulated data at future experiments - like the ones that will run at the High Luminosity Large HadronCollider (HL-LHC) - are expected to increase by orders of magnitude, increasing drastically thecomputational challenge. This expectation motivates the research for alternative deep learning-basedsimulation strategies.In this research we speed-up HEP detector simulations for the specific case of calorimeters usingGenerative Adversarial Networks (GANs) with a huge factor of over 150 000x compared to thestandard Monte Carlo simulations. This could only be achieved by designing smart convolutional 2Dnetwork architectures for generating 3D images representing the detector volume. Detailed physicsevaluation shows an accuracy similar to the Monte Carlo simulation.Furthermore, we quantize the data format for the neural network architecture (float32) with the IntelLow Precision Optimization tool (LPOT) to a reduced precision (int8) data format. This results in anadditional 1.8x speed-up on modern Intel hardware while maintaining the physics accuracy. Theseexcellent results consolidate the beneficial use of GANs for future fast detector simulations