Estimation of the Average Size of Crystals from a CMSMPR crystallizer by a neural network model

This work show the feasibility of using a neural network based model (NNM) to predict the dynamic behaviour of a K2SO4’s crystallizer. Firstly, an experimental investigation was carried out at pilot scale and a phenomenological model was implemented. Then, a NNM was trained and validated on the basis of data sets generated by the physical model. In operating continuous crystallizers the main objectives are to assure a high purity and a suitable size distribution of the crystal product. An efficient control strategy cannot be applied unless reliable methods for measuring the crystal size distribution (CSD) are available. Many attempts have already been made to measure the crystal size distribution on-line. However, even though some of the proposed measuring techniques have proved to work successfully on laboratory scale, their transfer to industrial scale still appears troublesome. An alternative approach is to develop an inferential technique for the estimation of the average size from a more easily measurable particle suspension characteristic. The main aim of this work is to investigate the feasibility of inferring the value of the average size of the crystal population from the magma density of fines, by using a dynamic neural network model (NNM) as estimator. The cooling crystallization of potassium sulphate from aqueous solutions has been chosen as case study. The work has been carried out in three steps: i) an experimental investigation on the performances of a continuous cooling crystallization runs, operating at steady and unsteady state; ii) the implementation of a phenomenological dynamic model and its validation on the basis of the obtained crystallization results; iii) the development of a neural network model based inferential estimator, trained and validated on data sets generated by the above mentioned phenomenological model.