Jussi Auvinen (IFT)

An exotic state of matter called quark-gluon plasma (QGP) is formed in high-energy heavy-ion collisions performed at Relativistic Heavy Ion Collider (RHIC) in USA and in the Large Hadron Collider (LHC) at CERN. Determining the properties of QGP is a challenging research topic requiring complex models with several parameters which have to be tuned to match the data. Bayesian statistical analysis is a method to obtain probability distributions for these model parameters, which provides not only the best-fit values but also the associated uncertainties. However, this posterior probability distribution has to be evaluated numerically which requires thousands of evaluations of the heavy-ion collisions model used in the analysis. As these models are computationally expensive, emulators need to be utilized. In this talk I present a Bayesian analysis with two layers of model emulation: First, neural networks are used to estimate model output for a single event, which allows us to evaluate hundreds of thousands of events with reasonable computational resources. Based on these single-event neural network computations, Gaussian process emulators are then trained to produce estimates for the event-averaged output which can be compared with the experimental data. This procedure allows us to obtain constraints on the temperature dependence of shear and bulk viscosities in QGP.