The aim of this work was to elaborate a predictive model of the mass transfer (water loss and solute gain) that occurs during dewatering and soaking by using neural network modelling. Two separate feedforward networks with one hidden layer were used (for water loss and solute gain respectively). Model validation was carried out on results obtained previously, which dealt with agar gel soaked in sucrose solution over a wide experimental range (temperature, 30-70 °C; solution concentration, 30-70 g sucrose/100g solution; time 0-500 min; agar concentration, 2-8%). The best results were obtained with three hidden neurons, which made it possible to predict mass transfer, with an accuracy at least as good as the experimental error, over the whole experimental range. The technological interest of such a model is related to a rapidity in simulation comparable to that of a traditional transfer function, a limited number of parameters and experimental data, and the fact that no preliminary assumption on the underlying mechanisms was needed.