In　order　to　predict　the　flux　of　BSA　solutions　under　the　different　operating　conditions　(transmembrane　pressure,　feed　concentration　and　temperature)　in　the　dead-end　microfiltration,　the　training　epochs,　correlative　coefficient　and　relative　absolute　error　were　used　as　three　predictive　criterions,　and　the　configurations　of　the　developed　three　layers　BP　and　RBF　neural　network　were　optimized　by　changing　the　interior　parameters　of　neural　networks.　The　result　showed　that,　in　the　experimental　rang,　an　optimal　configuration　of　the　available　BP　neural　network　is　3-9-1,　the　number　of　hidden　neurons　is　9,　the　learning　rate　is　0.05,　the　learning　function　is　traingdx,　the　transfer　function　is　logsig.　By　using　the　BP　neural　network,　the　obtained　average　relative　absolute　error　and　correlative　coefficient　is　2.37%,　0.9960,　respectively.　By　using　the　RBF　neural　network,　the　designable　function　of　the　network　for　an　optimal　configuration　of　the　available　RBF　neural　network　is　newrbe,　and　its　spread　is　400.　The　obtained　average　relative　absolute　error　and　correlative　coefficient　is　4.83%,　0.9870,　respectively.　Therefore　an　BP　neural　network　is　much　better　than　a　RBF　neural　network　in　the　study　of　predicting　the　flux　of　BSA　solutions　in　the　dead-end　microfiltration.