Computer codes used in support of nuclear industry are more and more complex, and consequently more and more CPU time consuming. We are here interested in such code, in the special case of functional output : the code output represents the evolutions of some physical parameters in time. Those last curves are functions from an interval $I \subset \R$ to $\R$, which will be preprocessed in order to cluster them in a few meaningful groups (clustering, or unsupervised classification). The aim of our work is the estimation of the convergence speed of clustering error estimates. After finding bounds on convergence speeds, we will illustrate this on an example with six distinct groups of curves.