Experimental results published in the literature between 1935 and 2000 were used to generate a working database of 558 loading capacity data for randomly dumped packed beds. The reported measurements were first used to review the accuracy of the few available predicting loading capacity correlations. The Billet and Schultes semiempirical correlation (Trans IChemE 77 (1999) p. 498) emerged as the best prediction method and is recommended for loading transition estimation, only when the constant C-SO of a given packing element is available. When such a model-dependent parameter is unavailable, an alternative and generalized neural network correlation is proposed to improve the broadness and accuracy in predicting the loading capacity for packed towers. A combination of five dimensionless groups, namely the liquid Reynolds (Re-L), Galileo (Ga-L) and Stokes (St(L)) numbers as well as the packing sphericity (phi) and one bed number (S-B) outlining the tower dimensions were used as inputs of the neural network correlation for the prediction of the loading capacity via the Lockhart-Martinelli parameter (chi),The correlation yielded an absolute average relative error of 21 % and a standard deviation of 19.9 %. Through a sensitivity analysis: the Stokes number in the liquid phase exhibits the strongest influence on the prediction while the liquid velocity, gas density and packing surface area are the leading physical properties defining the loading level.