Fast approximation of SUSY NLO cross-sections using Deep Learning

24 Jul 2018, 14:54
Plaza 1

Plaza 1


Sydney Otten (Radboud University Nijmegen and University of Amsterdam)


Although deep learning might appear as a magic black box one only needs to throw data at to receive a solution, the practical reality is always different and often difficult. This talk shall guide through an example for the process of constructing an AI for predicting a physical quantity, namely the pMSSM-19 NLO electroweakino production cross-section at the LHC for 13 TeV. Naively, one could assume that this is merely a simple regression task but as I will demonstrate, classifiers, active learning and feature engineering including an injection of deeper expert knowledge played a crucial role for solving this task and will, if not necessary, at least be beneficial for many other tasks. While the available Monte Carlo methods take several minutes per cross-section, the resulting AI is able to deliver $\approx\,10^5$ NLO cross-sections per second with an average error of about $0.1\,\%$ and a maximum error below uncertainties.

Parallel Session Precision Calculations and MC tools

Primary authors

Sydney Otten (Radboud University Nijmegen and University of Amsterdam) Krzysztof Rolbiecki (University of Warsaw) Roberto Ruiz De Austri Jong Soo Kim Jamie Tattersall (RWTH Aachen) Sascha Caron (Nikhef National institute for subatomic physics (NL))

Presentation materials