INTEREST OF NEURAL NETWORKS FOR THE OPTIMIZATION OF THE CROSS-FLOW FILTRATION PROCESS

In order to build up a model representing the effect of transmembrane pressure and crossflow velocity on crossflow filtration results at qua si-steady state, an approach based on neural networks is proposed. For filtrations of various products (raw cane sugar remelt, natural gum s olution) on different membranes (micro- and ultrafiltration) with or w ithout co-current permeate flow, the modelling of both permeate flux a nd retention rate could be obtained after only five experimental trial s. Compared to more classical modelling techniques, the neural network s were showed to be sometimes better suited and are useful when the ef fects of hydrodynamical conditions on filtration results are strongly nonlinear. Thanks to established models, it was possible to determine with a good safety margin, an optimum region in every case studied.