With the climate change emergency, pollutant and fuel consumption reductions are now a priority for aircraft industries. In combustion chambers, the chemistry and soot modeling are critical to correctly quantify engines soot particles and greenhouse gases emissions. This thesis aimed at improving aircraft numerical pollutant tools, in terms of computational cost and prediction level, for engines high fidelity simulations. It was achieved by enhancing chemistry reduction tools, allowing to predict CO emissions of an aircraft engines at affordable cost for the industry. Next, a novel closure model for unresolved terms in the LES filtered transport equations is developed, based on neural networks (NN), to propose a better flame modeling. Then, an original soot model for engine high fidelity simulations is presented, also based on NN. This new model is applied to a one-dimensional premixed sooted flame, and finally to an industrial combustion chamber LES with measured soot comparison.