Abstract: The objective of this study was to establish a BP neural network to predict formononetin retention rate in the process of ultrafiltration for fibrillary rhizome herbs.Using the membrane pore size,operating pressure and temperature as input variables,formononetin retention rate as output variable,the BP neural network was established after network parameters optimized using L-M method.Furthermore,the performance and applicability of model were evaluated.The optimal process condition and the effects of different ultrafiltration conditions on formononetin retention rate was then investigated.The average error rate of Hedysari Radix and Astragali Radix  was 1.78% and 1.92%,respectively.The optimal ultrafiltration process was as follows:membrane pore size 100 nm,operating pressure 0.15 Mpa,temperature 45 ℃.The influence of formononetin retention rate was temperature＞membrane pore size＞operating pressure.The results showed that the BP neural network had good practical value because of higher accuracy and better applicability.The established prediction model can avoid optimizing ultrafiltration technology of similar fibrillary rhizome herbs repeatedly.

Key words: Hedysari Radix, Astragali Radix, formononetin, enzymolysis, ultrafiltration, BP neural network, prediction model