This report summarizes the main achievements of the first year’s effort of development of new crystallization technologies for improved crystal size and shape control during the crystallization process. The successful crystallization process and system design requires an interdisciplinary effort, which ranges from population balance model (PBM) development of the system concept, through efficient implementation of model equations to soft-sensor development, which is required for the model predictive control (MPC) design as well. This report contains practically the first steps of these interdisciplinary developments, which already crossed each-other in certain points, but it will consist organic parts of the final, integrated system.
The system concept analyzed consists in implication of wet-milling during the crystallization, which is supposed to widen the achievable crystal size domain, and this is applied as an external equipment linked to the crystallizer by a recirculation line. The recirculation flowrate also serves as an important design parameter that can be optimized. In order to make the optimization feasible, the nonlinear equation system must be implemented efficiently. In this work the implication of the graphical processing units (GPU) is employed to speed up the solution of the high-resolution finite volume method (HR-FVM). The GPU ensured roughly 100% speedup for the simulation of integrated batch-crystallization external wet mill system for the 1D case. Also, a generic, portable crystallization modeling platform for 1D and 2D batch and continuous crystallization systems have been developed. The dynamic optimization of the integrated crystallizer – wet mill system revealed an unexpected, sequential operation: the primary nucleons created in the crystallizer are transferred to the wet mill, where the (very uncertain) population is milled down to the minimal size, during which in the crystallizer complete dissolution occurs. Then, the seed crystals are fed back to the crystallizer dynamically in controlled manner, which is able to achieve broad variety of CSD shapes.
A forward transformation, with great accent on real time applicability, is developed for 2D rod-like crystals to simulate the most probable chord length distribution (CLD) and aspect ratio distribution (ARD) that would be measured. In order to speed up the CLD and ARD calculation of arbitrary 2D CSDs artificial neural networks (ANN) are implied. This will be applied in a CLD based N-MPC of crystal size and shape, which is subject of further development.