Looking for inflationary features – the Bayesian way

The seeds of the Universe’s large scale structure are widely thought to have been generated during a period of cosmic inflation. Simple, generic models of inflation lead to a smooth, almost scale-invariant spectrum of primordial curvature perturbations. However, this prediction relies on a set of physical assumptions, which, if broken, could lead to a power spectrum with strong scale dependence in the form of, e.g., cutoffs spikes or modulations – commonly referred to as inflationary “features”. The search for features is complicated by increased numerical precision requirements and typically non-trivial shapes of the likelihood function – taking conventional random sampling-based analysis methods to their limits. I will introduce a new, efficient machine-learning approach based on Gaussian Process Regression and Bayesian Optimisation. Applying it to search for inflationary features in Planck data, I will demonstrate its effectiveness and discuss further applications, such as the calculation of Bayesian evidences for model selection.