# Dortmund 2021 – wissenschaftliches Programm

## Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

# T: Fachverband Teilchenphysik

## T 38: Data analysis, information technology II

### T 38.3: Vortrag

### Dienstag, 16. März 2021, 16:30–16:45, Tm

**Using a conditional Invertible Neural Network to determine the parameters of ultra-high-energy cosmic ray sources** — Teresa Bister, Martin Erdmann, and •Josina Schulte — III. Physikalisches Institut A, RWTH Aachen University

The usage of modern machine learning techniques is growing with immense speed and new advanced methods for physics analysis are developed. Often, the task is to estimate the physical parameters of a model, using only a set of measurements without the possibility to formulate an explicit inverse function. Traditionally, the full posterior parameter distributions can be approximated using the Markov Chain Monte Carlo (MCMC) method. This is useful to uncover correlations and to estimate the parameter uncertainties. However, posterior sampling is generally computationally expensive and slow. A new alternative technique, based on Deep Learning, is a so-called conditional Invertible Neural Network. Here, the network implicitly learns the posterior distributions from a large set of training data, whereby a specific loss function ensures convergence. The basic functionality and possibilites of this technique, applied on an example from astroparticle physics, will be presented in this talk. Here, the observable is the binned energy spectrum of ultra-high-energy cosmic rays on Earth. The free parameters of the astrophysical model define the acceleration process at the origin sites, which are currently still unknown. We show that it is possible to find posterior distributions of the free model parameters with the network and compare them to the ones using the classic MCMC method.