A new dynamic optimization technique presented combines a neural network model with a universal dynamic matrix control (UDMC) algorithm. This technique utilizes a nonlinear-model-predictive control technique for on-line optimization and feedback control by using a dynamic neural net model. This approach offers two important advantages over conventional UDMC. One is that a dynamic neural net model can be developed from process data and used for optimization calculations, thus achieving optimization without a first principle model. This neural-network-based optimization approach also produces good performance even with process-model mismatch. The other is that our neural-net-model-based UDMC algorithm greatly reduces the computation time required for the nonlinear dynamic matrix used for the successive quadratic programming algorithm. The development of this technique also involved an analysis of the effect of network structure on dynamic optimization. A state-space-based neural network model which utilizes a priori process knowledge is best suited for optimization calculations. Advantages of this technique are illustrated by simulation for two chemical processes.