Multiphysics modeling is a powerful tool for analyzing nuclear reactors, but the uncertainty propagation between disciplines is often disregarded. This PhD thesis proposes innovative approaches to improve the accuracy of multiphysics modeling by accounting for these uncertainties. The primary goal is to propose optimal modeling approaches tailored to diverse accuracy requirements. This information is of prime interest to researchers and industry professionals involved in the development and utilization of multiphysics models. Specifically, the thesis will assess various uncertainty propagation techniques applicable to multiphysics simulations. This involves exploring surrogate modeling through avenues like reduced-order modeling and polynomial chaos expansion. The goal is to identify and categorize input parameters with the most significant impact on system outputs, irrespective of their physical domain. Subsequently, uncertainty propagation will be executed using two core modeling types: a ‘high-fidelity’ model based on the CEA's reference simulation tools and a ‘best-estimate’ model accounting for the "industrial" objective of the calculations). The similarities and differences between these approaches will be analyzed to assess model biases. These uncertainty evaluations employing the above methods will be tested on an extensive set of experiments performed in SEFOR, a sodium-cooled fast reactor, representing a diverse range of experimental data for various reactor conditions.