Speaker
Description
Exploring the use of Bayesian neural networks for the purpose of evaluating supersymmetric cross sections at next-to-leading order with reliable uncertainty estimates. Calculating cross sections is a central part of the search for new physics beyond the Standard Model, however, it is a time consuming endeavour. Machine learning methods can speed up the production process by learning the relationships between physical parameters and cross sections, however, the uncertainty on their predictions is for most methods unavailable. A variational Bayesian neural network was trained on cross section data produced by Prospino 2.1 for electroweakino pair production in proton-proton collisions. To improve its performance deterministic pretraining was implemented and various annealing schemes were tested. We found that with the additional implementations the Bayesian neural network consistently produced reliable uncertainties and favourable predictions with relatively short run-times.