Influence of selected process parameters on attrition intensity in DTM type crystallizers with a jet–pump – a general neural network’s approach
Advancing the chemical engineering fundamentals
Crystallization (T2-9P)
Keywords: sodium chloride crystals, attrition degree, coefficient of variation, liquid jet–pump, neural network model
For properly designed industrial–scale mass crystallization process a main source of nuclei is a complex secondary nucleation, based mainly on attrition, breakage and abrasion phenomena. In the most common crystallizer constructions with internal circulation of suspension this desirable (if strictly controlled) effect results from the mechanical collisions between: crystal–crystal, crystal–agitator (or pump rotor) and crystal–apparatus wall (or/and other interior equipment). However, in case of too intensive attrition action an excessive number of nuclei arises, causing difficulties in process control and making creation of the product of desirable crystal size distribution impossible. Facing the complex interrelations between hydrodynamic and constructional factors the exact prediction of attrition action in the assumed process environment is an important engineering challenge. Application of original constructions of crystallizer with a liquid jet–pump creates new possibilities of providing stable and intensive enough internal circulation of suspension inside the vessel simultaneously reducing the undesirable excessive attrition effects.
Experimental tests of the attrition resistances in NaCl crystal population (hardness via Vickers method 16 – 24 MN/m2) of initial mean size within the range of Lm = 0.441 – 1.52 mm were performed. Crystal volumetric concentration in the suspension was adjusted in the j = 2 – 10 % range, while the suspension residence time in the crystallizer was changed in the t = 900 – 7200 s range. Two different constructions of laboratory–scale crystallizer with a jet–pump were taken under consideration: DTM MSMPR (Draft Tube Magma Mixed Suspension Mixed Product Removal crystallizer) and a more complex design with internal hydraulic classification system – DTM MSCPR (DTM Mixed Suspension Classified Product Removal crystallizer). Both laboratory crystallizers, loaded with the crystal suspension of possibly narrow, closely restricted crystal size distribution (saturated solution, r = 1198 kg/m3, T = 298 K) were working under steady state mode through the selected residence time. The resulting, final crystal size distribution was determined with Laser Particle Size Analyzer COULTER LS–230. For the assumed values of process parameters each measurement was repeated twice (repeatability test).
Taking under consideration complex entirety of analyzed phenomenon (e.g. influence of hydrodynamics in macro- and microscale) a feedforward multilayer artificial neural network was used for creating a numerical model of this process based on the experimental dataset (60 input–output vectors). The net was composed of three inputs (representing: mean residence time of suspension, crystals volumetric concentration in the suspension and initial mean size of crystals) and six output neurons, representing: mean size of crystal population after the process, attrition degree (defined as a ratio of the difference of mean sizes – before and after the test – to the initial mean size) and final value of coefficient of variation (CV) – a set for DTM MSMPR and a set for DTM MSCPR crystallizer, respectively. The 30 neural network configurations with 4 – 24 hidden neurons arranged with 1 – 2 hidden layers were a subject of statistical tests (RMSD – Root Mean Square Deviation). An optimal configuration proved to be a 3–6–6 structure, trained for 8600 iterations with learning rate set as 0.1. Its statistical accuracy was as follows: for DTM MSMPR apparatus RMSD for attrition degree was 0.8810 %, for Lm – 0.0145 mm and for CV – 0.9847 %. For DTM MSCPR apparatus RMSD for attrition degree was 0.5724 %, for Lm – 0.0157 mm and for CV – 0.9164 %.
The resulting general, numerical neural model enables one to predict (and then eventually control) with the significant accuracy attrition intensity for NaCl crystals in various technological conditions, including: mean size of the process magma, its volumetric concentration and mean residence time in the apparatus. Based on experimental data, thus devoid of any simplifying assumptions, the model can correctly render all possible hidden, strongly nonlinear intrinsic interrelations and feedbacks between these three process factors, imposing simultaneously the hydraulic regime resulting from the selected geometrical arrangement of the vessel’s interior of two different constructional solutions (mixed and classified product removal). The experimental and simulated data, however linked strictly quantitatively only with the laboratory–scale crystallizers used, can provide one with at least qualitative information about attrition behavior of NaCl crystal suspension in a large–scale liquid jet–pump crystallizer construction of diversified internal hydraulic regimes, making the comparative study as well as optimal process conditions location possible.
See the full pdf manuscript of the abstract.
Presented Wednesday 19, 13:30 to 15:00, in session Crystallization (T2-9P).
