Simulation of Extensive Air Showers with Deep Neural Networks - PowerPoint PPT Presentation

Simulation of Extensive Air Showers with Deep Neural Networks Marcel Köpke Auger Youngster Meeting (2019) INSTITUTE FOR NUCLEAR PHYSICS (IKP), FACULTY OF PHYSICS KARLSRUHE INSTITUTE OF TECHNOLOGY (KIT) www.kit.edu KIT – The Research University in the Helmholtz Association

CORSIKA 7 [1] Extensive air shower Monte Carlo simulation framework Different types of interaction models (EPOS-LHC, QGSJET, SIBYLL, ...) 1 TeV Proton 1 TeV Iron 10 TeV Proton 10 TeV Iron Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 2 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Motivation The time complexity of CORSIKA 7 simulations rises approximately linearly with the primary particle energy Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 3 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Thinning Reduces (effective) particle content by particle-aggregation Preserves shower properties to leading order Reduces shower-to-shower fluctuations Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 4 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Why Neural Networks? Can run on specialized hardware (GPU / TPU) Automatic parallelization (TensorFlow) Automatic reduction to essential features Training can fix meta-parameters Adjustable accuracy possible Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 5 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Neural Networks Combination of linear and non-linear functions Training via loss function / metric on data pairs Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 6 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Generative Adversarial Neural Network (GAN) Train discriminator on real (1) and generated (0) data Train generator to outsmart the discriminator Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 7 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Training: Discriminator (Part 1) Train discriminator on real (1) and generated (0) data Train generator to outsmart the discriminator Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 8 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Training: Sampling Train discriminator on real (1) and generated (0) data Train generator to outsmart the discriminator Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 9 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Training: Discriminator (Part 2) Train discriminator on real (1) and generated (0) data Train generator to outsmart the discriminator Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 10 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Training: Generator Train discriminator on real (1) and generated (0) data Train generator to outsmart the discriminator Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 11 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Training: Result Train discriminator on real (1) and generated (0) data Train generator to outsmart the discriminator Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 12 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Training: Result Train discriminator on real (1) and generated (0) data Train generator to outsmart the discriminator Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 13 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Training: Result Train discriminator on real (1) and generated (0) data Train generator to outsmart the discriminator Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 14 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Training: Result Train discriminator on real (1) and generated (0) data Train generator to outsmart the discriminator Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 15 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

First Test (CONEX) CONEX: Hybrid Extenisve Air Shower Simulation first: Monte Carlo until energy threshold (3D) – then: cascade equation solver (1D) – provides longitudinal profile only – runtime: seconds – minutes – Configuration: E = 1E17 ... 1E19 eV – Zenith = 0 ... 65 deg – Azimuth = -180 ... 180 deg – Generated ~187k datapoints Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 16 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Shower-to-Shower Fluctuations Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 17 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Shower-to-Shower Fluctuations Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 18 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Shower-to-Shower Fluctuations Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 19 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

CONEX vs. GAN Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 20 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

CONEX vs. GAN Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 21 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

CONEX vs. GAN Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 22 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Shower Library Shower library required for analyses and model training Trained model = effective compression of shower library Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 23 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

What‘s next? Fix it (oversampling, architecture, ...) (Meta)parameter tests Test adversarial vulnerability Template matching/reconstruction Refining with data Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 24 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Backup Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 25 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Fast Implicit Simulation Heuristic (FISH) Autoencoder with Adversarial Metric Simulation Input (SI) can be extended with meta-parameters Discriminator can be refined with real measurements Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 26 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

Adjustable Accuracy [2] ResNet Translate to ordinary differential equation (ODE) Solve with standard ODE solver Adapt solver accuracy on the fly (training: high, inference: low) Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 27 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)

References Title picture: Karlsruhe Castle - Meph666 [CC BY-SA 3.0] https://commons.wikimedia.org/wiki/File:Karlsruhe-Schloss-meph666- 2005-Apr-22.jpg Backup picture: Photo by Anthony from Pexels [1] CORSIKA 7: https://www.ikp.kit.edu/corsika/ [2] „Neural Ordinary Differential Equations“ - Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, David Duvenaud – arXiv: 1806.07366 Marcel Köpke: Simulation of Extensive Air Showers with 23.09.2019 28 Institute for Nuclear Physics (IKP), Faculty of Physics Deep Neural Networks Karlsruhe Institute of Technology (KIT)