In the framework of risk assessment in nuclear accident analysis, best-estimate computer codes, associated to a probabilistic modeling of the uncertain input variables, are used to estimate safety margins. A first step in such uncertainty quantification studies is often to identify the critical configurations (or penalizing, in the sense of a prescribed safety margin) of several input parameters (called “scenario inputs”), under the uncertainty on the other input parameters. However, the large CPU-time cost of most of the computer codes used in nuclear engineering, as the ones related to thermal-hydraulic accident scenario simulations, involve to develop highly efficient strategies. This work focuses on machine learning algorithms by the way of the metamodel-based approach (i.e., a mathematical model which is fitted on a small-size sample of simulations). To achieve it with a very large number of inputs, a specific and original methodology, called ICSCREAM (Identification of penalizing Configurations using SCREening And Metamodel), is proposed. The screening of influential inputs is based on an advanced global sensitivity analysis tool (HSIC importance measures). A Gaussian process metamodel is then sequentially built and used to estimate, within a Bayesian framework, the conditional probabilities of exceeding a high-level threshold, according to the scenario inputs. The efficiency of this methodology is illustrated on two high-dimensional (around a hundred inputs) thermal-hydraulic industrial cases simulating an accident of primary coolant loss in a pressurized water reactor. For both use cases, the study focuses on the peak cladding temperature (PCT) and critical configurations are defined by exceeding the 90%-quantile of PCT. In both cases, the ICSCREAM methodology allows to estimate, by using only around one thousand of code simulations, the impact of the scenario inputs and their critical areas of values.