Uncertainty quantification and variability analysis are two domains of interest when looking at the efficiency of HPEM sources. Vircator is known to be a low efficiency high power microwave source subject to several generally volatile phenomena such as plasma expansion and shot-to-shot variability. In this study, a computationally low cost framework combining the Extreme Value Theory (EVT) and the Generalised Design of Experiments is proposed in order to study the peak power distribution of a Vircator obtained with a surrogate model. Following the pre-screening of random variables, the optimised parameters are introduced in 2.5D and 3D simulation tools, namely XOOPIC and CST-PS. It has been confirmed that the peak power output can reach a 40% increase. This shows that the EVT proves to be successful in classifying and quantifying random variables to influence the distribution tails.