The influence of the properties of commercial powders on the densification during their packing, compaction and sintering process is still not understood in detail. With regard to the sintering process, neither the well-known sintering equation for the first sintering step nor the relation between the density and grain size at the final sintering step can describe the influence of powder characteristics on its densification behaviour. For improving the sintered density of a ceramic powder, it is known to be crucial to start with a highly dense and homogeneous green body. Therefore, the powder has to fulfil different requirements such as being agglomerate free, reasonably spherical and having a narrow size distribution (but not mono-dispersed). The aim of this work is to develop a better understanding of the relation between the powder properties and the densification behaviour during the packing, compaction and sintering process, of commercial, micron sized, metallic and ceramic powders. Another aim of this work is to evaluate if prediction of the packed, green and sintered densities based only on the known powder characteristics is possible via a neural network approach. The presented results show that a well learnt neural network is a useful tool for the prediction of green and sintered densities of granulated alumina powder produced either by milling (Bayer process) or by chemical processes. Moreover, the simulated influences of characteristics, on the green and sintered densifications, fit well literature models behaviours and intrinsic properties of such powders. Concerning the green densification, Bayer powders are denser for coarser particles and/or a broader size distribution. Relating to the chemically produced powders, those tendencies are inversed, due a stronger agglomeration with a broader size distribution and coarser particles. Regarding the sintered density, the neuronal approach highlights a better sinterability for finer powders. Limits of the artificial neural network tool are emphasized with its application to metallic powders: the learning stage seems to be primeval and simulated results are to be analysed and interpreted with care and inside the validity domain of each specific artificial neural network.