A novel neural network architecture is presented for dynamic process modeling, using stirred cell microfiltration of bentonite suspensions as a model system. Unlike previous studies that include time explicitly as a network input and have a single output at that time, the network architecture presented contains the process variables as inputs and many outputs representing the output (filtrate flux in this case) at different selected times. The network is shown to represent the stirred cell microfiltration of bentonite suspensions over a range of pressures (0.2-1.5bar), initial concentrations (0.5-2.0g/L), stirrer tip speeds (0.04-0.17m/s), membrane resistances (3.09x1010-6.85x1010m-1), pH values (2.5-10.4), and temperatures (20 degrees-24 degrees C) with good accuracy (R2=0.91 on network test data). With this network architecture, it becomes easy to track the time dependence of the relative effect of the various process parameters on the system output. Thus, for example, the network weights show that the effect of stirring rate on flux increases as time progresses, while the opposite effect is seen for membrane resistance, as expected.