Abstract: Composite model for bioprocesses is presented by combining mass balance cquations of bioreactors and with neural networks which serve as estimators of unmeasutedprocess kinetic parameters. The novel neuron transfer functions of Monod saturation andsubstiate inhibition form are developed. They are very useful in modeling fed-batch cellgrowth and other bioprocesses. Based on Hebbian learning rule, a fast learning algorithm(FLA ) with error back-propagation at one time is proposed for the training of feedforward neural networks by introducing supervisory learning signal. Composit model is applied for theprediction and estimation of state variables , specific cell growth rate and specific substrateconsumption rate. Simulation results show that composite model is able to interpolate and extrapolate much more accurately. It is easier to analyze and interpret , requires fewer trainingexamples and has faster convergence speed ,stronger noise immunity and robustness.