This thesis aims at detecting large-scale turbulent structures in incompressible Rayleigh-Taylor mixing layers at low Atwood number. Various statistical quantities conditioned by structure presence have been obtained and it is now possible to compare them with results from two-structure statistical turbulent models such as the 2SFK model developed at CEA. In order to produce direct numerical simulations of the turbulent mixing, a three-dimensional, incompressible, variable-density numerical code was developed. This code is parallelized in the three directions. Several structure detection methods have been designed and tested. Although all these methods are of interest, only the most efficient with respect to our detection criteria has been retained for simulations at high resolution (over a billion cells, 1024^3). A time filtering of vertical velocity is used in this method to: 1) correct distortions due to stagnation points and recirculation zones in the flow, 2) minimize small-scale turbulence effects and better highlight large-scales, 3) introduce a memory effect in order to extend bimodality of the detection field from the external laminar zones up to the centre of the turbulent mixing zone. Several direct numerical simulations at 1024^3 have been achieved. Results support those obtained with two-structure 2SFK model and justify further studies for its validation.
