In spark-ignited direct-injected engines, the formation of fuel pools on the piston is one of the major promoters of unburnt hydrocarbons and soot: in order to comply with the increasingly stringent emission regulations (EU6 and forthcoming), it is therefore necessary to limit fuel deposit formation. The combined use of advanced experimental techniques and detailed 3D-CFD simulations can help to understand the mechanisms driving fuel pool formation. In the paper, a combined experimental and numerical characterization of pool formation in a GDI engine is carried out to investigate and understand the complex interplay of all the mentioned factors. In particular, a low-load low-rpm engine operation is investigated for different ignition phasing, and the impact of both fuel formulation and instantaneous piston temperature variations in the CFD analyses are evaluated. The investigated engine operation shows some interesting features which are suited to deeply investigate the interplay between fuel film formation, mixing and soot. In particular, the relatively low wall temperature and low injection pressure allow the fuel to form deposits and then slowly evaporate, with possible presence of liquid fuel at the time of ignition. The simultaneous presence of slow fuel evaporation, reduced turbulence and presence of liquid fuel leads to the formation of extremely rich mixture pockets (with equivalence ratios well above 5) which are the major promoters for soot inception.Four different start of injection (hereafter SOI) values are analyzed, for which tailpipe Soot concentration measurements are available. For one SOI value, two different injection profiles are also evaluated. In particular, the analyses focus on the formation of fuel pads on the combustion chamber walls and on the mixture stratification, and a correlation between these two factors and the tailpipe soot level is found.The proposed methodology proves to be able to capture the Soot trend for the different SOI values without simulating the combustion process; it is therefore promising since it avoids the need for a dedicated calibration of the combustion model parameters and provides reasonable results (at least in terms of trends) with limited computational resources.