Reliable estimation of the melt index (MI) is crucial in the quality control of practical propylene polymerization (PP) processes. In this paper, a novel predictive neural network system, combining the particle swarm optimization (PSO) algorithm and radial-basis function neural networks (RBFN), is presented to infer MI from real PP process variables, where the PSO algorithm dynamically constructs the RBFN structure and parameters and a new adaptive PSO (APSO) algorithm, which adjusts the algorithm behavior based on evolution information of swarms, further accelerates the convergence speed. Principle component analysis is applied to select the most relevant process features and to reduce the number of input variables in the model. A detailed comparison between PSO, APSO and the gradient descent algorithm is carried out using historical data from a real plant.