An experimental verification and validation of the neural-network rate-function (NNRF) approach to modeling dynamic systems is provided. The NNRF modeling scheme utilizes the designed experimental data to build the interested states of the dynamic system as a function of the related recurrent states and the external inputs by feedforward artificial neural networks (FANNs). Once these FANNs are trained, they can be put back into the original differential equations, thereby turning out this combined NNRF-differential-equation model. The experimental system for demonstrating the applicability of the NNRF modeling approach is the liquid phase cultivation of Monascus anka. In a 5 l batch reactor system, the controlled inputs are the cultivating temperature and the dissolved oxygen; the desired outputs are the glucoamylase activity and the concentration of red pigment. Based on the developed sequential pseudo-uniform design (SPUD) method, 11 batches are adequate to provide the identified NNRF model with sufficient information. Facing the noisy inputs and outputs, the identified NNRF demonstrates its predictive capability and could be applied to determine the optimal operating conditions for the submerged cultivation of M. anka.