A recurrent backpropagation neural network (RBPN) was proposed for the on-line adaptive pH control of penicillin acylase fermentation with Arthrobacter viscosus. It was observed that both enzyme activity and cell growth are rather sensitive to changes in pH. Hence, the control of pH during batch fermentation is a very important issue. RBPN was chosen as the controller model for its superior ability in long-term identification. The transfer function x/(1 + x) proposed previously was used with this RBPN controller. The output node of this network controller was the predicted flow rate for the next control time interval. Initial pump rate and base/acid concentrations were both important factors affecting the control performance. To enhance the effective on-line learning of this network, a moving-window type of training data was supplied to train the network. In conclusion, the pH was well controlled and a maximum optical density of 6.7 was achieved as well. Therefore, a test of the RBPN controller from the pH control of this fermentation was successfully performed.