The evaluation of nuclear data with the production of international files integrated into an international library such as the European library (JEFF) is of major importance for calculating current and future nuclear systems and reactors. Understanding and mastering the uncertainties related to these nuclear data is a particularly delicate task, which requires the use of advanced Bayesian inference techniques. The objective of this PhD thesis is to develop a BEPU (Best Estimate Plus Uncertainty) approach of the CEA neutronics codes which is holistic in the sense that all the known sources of uncertainty are taken into account (nuclear data, geometry and material description data, model approximations,...) when solving the neutron transport equation. In oder to account for these aleatory and epistemic uncertainties, we will use both a standard Bayesian framework and recent machine-learning methods (Deep Learning). In particular, this PhD thesis will focus on the difficult task of assimilating data from integral (critical and post-irradiation) measurements such as those available in the IRPhE international database. This work is essential for the validation of the new JEFF4 nuclear data library.