In the framework of risk assessment in nuclear accident analysis, best-estimate computer codes are used to estimate safety margins. Several inputs of the code can be uncertain, due to a lack of knowledge but also to the particular choice of accidental scenario being considered. The objective of this work is to identify the most penalizing (or critical) configurations (corresponding to extreme values of the code output) of several input parameters (called ``scenario inputs''), independently of the uncertainty of the other input parameters. However, complex computer codes, as the ones used in thermal-hydraulic accident scenario simulations, are often too CPU-time expensive to be directly used to perform these studies. A solution consists in fitting the code output by a metamodel, built from a reduced number of code simulations. When the number of input parameters is very large (e.g., around a hundred here), the metamodel building remains a challenge. To overcome this, we propose a methodology, called ICSCREAM (Identification of penalizing Configurations using SCREening And Metamodel), based on screening techniques and Gaussian process (Gp) metamodeling. The efficiency of this methodology is illustrated on a thermal-hydraulic industrial case simulating an accident of primary coolant loss in a Pressurized Water Reactor. This use-case includes 97 uncertain inputs, two scenario inputs to be penalized and 500 code simulations for the learning database. The study focuses on the peak cladding temperature (PCT) and critical configurations are defined by exceeding the 90%-quantile of PCT. For the screening step, statistical tests of independence based on the Hilbert-Schmidt independence criterion are used for global and target sensitivity analyses. They allow a significant reduction of inputs (from 97 to 20) and a ranking of these influential inputs by order of influence. Then, a Gp metamodel is sequentially built to reach a satisfactory predictivity of 82% of explained PTC variance, and a high capacity of identifying PTC critical areas (94% of good ranking rate above the threshold). Finally, the Gp is used to estimate, within a Bayesian framework, the conditional probabilities of exceeding the threshold, according to the two scenario inputs. The analysis reveals the strong interaction of the two scenario inputs in the occurrence of critical configurations, worst cases corresponding to medium values of both inputs.