A limit surface is an n-dimensional surface describing the nuclear plant status as function of selected plant parameters identifying the boundaries between failed and safe conditions for the reactor core. A limit surface can be used for the dynamic probabilistic risk assessment (PRA) analysis. In this paper, machine learning algorithms are applied to determine a limit surface for a PWR small break loss of coolant accident (LOCA). Basic ideas of the algorithms are similar to those of the RAVEN statistical analysis code. As a first step, a set of training simulations is run to construct a reduced order model for determining an approximated limit surface. The reduced order model is then used to predict where further exploration of the input space could be most informative. The new observations are used to update the reduced order model. This kind of adaptive samplings is iterated until convergence is obtained. A case study was carried out for the RELAP5 analysis of ROSA/LSTF small break LOCA tests with intentional secondary-side depressurizations. The limit surface was investigated for the peak cladding temperatures as a function of the break size and the depressurization timing. Several sampling methods were compared from the viewpoint of convergence to the limit surface. It was confirmed that using adaptive machine learning techniques allowed a remarkable reduction of the time required for the accurate limit surface determination.