Seppo Palosaari

Simulation of membrane separation by neural networks

Abstract A method for the simulation of membrane processes by using neural networks was developed. The neural network model is used for obtaining an estimation of permeate flux and rejection over the entire range of the process variables, i.e. pressure, concentration of solute, temperature and superficial flow velocity. Permeate flux and rejection are dependent on the process variables. This dependence has to be known in the flowsheet simulation of membrane processes. Traditionally it has been described by experimental polynomial correlation or by equations based on a mass transfer model. Neural networks offer the advantage of being easy to use. Furthermore, the computing time in the design of a membrane separation plant is shorter as compared with the calculation based on mass transfer models.

Mixing and crystallization in suspensions

The influence of mixing on suspension crystallization is described from the starting point of a fully mixed system with mixed product withdrawal. An imperfect suspension, caused by less intensive mixing, is then introduced and the effects of imperfect suspension on suspension density distribution along the height and on crystal size distribution in the product are shown both theoretically and based on available experimental results.

Calculation of permeate flux and rejection in simulation of ultrafiltration and reverse osmosis processes

Abstract A method for calculation of ultrafiltration and reverse osmosis processes has been developed to enable the calculation of permeate flux and rejection at different pressures and concentrations. The method has been combined with a process simulation program which calculates all streams of the process. Permeate flux and rejection in membrane processes are dependent on pressure and concentration of the solution. This dependence has to be known in the simulation of these processes. Normally a considerably large number of time consuming experiments have to be made to find the dependence of permeate flux and rejection on pressure and concentration. In this work the minimum experimental information was determined to enable the calculation of permeate flux and rejection. Two calculation procedures, one based on the finely porous model and the other on the statistical-mechanical model, were used. In these models permeate flux and rejection are described by four quantities. The values of these quantities are constant for each separation system having the same solution, membrane and temperature. A method of obtaining the values of these quantities from experiments is described and the minimum experimental information was determined to enable the calculation of permeate flux and rejection over the entire operating range of the process.

Application of neural networks for system identification of an adsorption column

This work illustrates the use of neural networks for system identification of the dynamics of a distributed parameter system, an adsorption column for waste-water treatment of water containing toxic chemicals. System identification of this process is done from simulated data for this work. The inputs to the networks include the state of the column at a given point in time and the system input, the velocity. The network predicts the change in the state over a period of time based on these inputs. Recurrent networks were found to be capable of simulating the whole operation of the column from an initial state of zero concentrations throughout the column, and thus predicting the complete breakthrough curves. The feasibility of system identification of this process has been established using synthetic noisy data, which indicates that the same can be performed from experimental data when all the required measurements are available.

Application of CFD simulation to suspension crystallization—factors affecting size-dependent classification

Abstract A size-dependent classification function, which is a parameter used to characterise crystallization in imperfectly mixed suspensions, was simulated using CFD. The CFX4.2 program was employed to simulate the local particle-size distribution in a multifluid flow using the k−e turbulence model with respect to interface transfer between the particles and the solution. The problem was studied three dimensionally. A sliding grid technique was used to simulate the transient flow in the mixing tank. The size-dependent classification was calculated based on the local particle-size distribution. The factors that affect the classification function are discussed.

Crystallization kinetics of potassium sulfate in an MSMPR stirred crystallizer

Abstract Experimental work on a potassium sulfate-water system was carried out using ten and fifty liter crystallizers. Different impeller velocities and suspension densities were used. The crystal size distribution was determined over the range from 0.1 μm to the largest crystals, which have been produced in the crystallizers, by the combination of Coulter LS-130 light scattering laser and by Vidas image analyzer results. Experimental evidence from continuous crystallizers frequently shows, at least for small crystals, deviation from the McCabe ΔL law. In this case, the estimation of kinetics of both nucleation and growth rate becomes more complicated. In this work the crystal size distribution was determined experimentally. The relation between growth rate and particle size is investigated. The methods of estimation of kinetics for industrial use are discussed. The experimental data of population density distribution was fitted directly by the three-parameter model presented by Mydlarz and Jones for a steady state MSMPR crystallizer. Then the relation between growth rate and particle size was calculated by the corresponding three-parameter growth rate model. The relation between growth rate and particle size shows that the apparent crystal growth rate increases linearly with the crystal size when the crystal size is smaller than about 10 μm, is strongly size-dependent when the crystal size is between 10–700 μm, and is size-independent when the crystal size is greater than 700 μm. A mechanism of growth rate dispersion is suggested.

Use of Neural Networks in the Simulation and Optimization of Pressure Swing Adsorption Processes

In this work a method for the simulation and optimization of a pressure swing adsorption process for the separation of nitrogen from air by using neural networks was developed. The model is used to obtain a prediction for the process performance, namely, the specific product and yield, over a wide range of operating conditions. These results are compared with the predictions from a mass tranfer model, and a very good agreement is found. The network developed is also used to minimize a cost objective function, and it is shown that it can easily be used in process optimization and/or control.

Flowsheet simulation of ultrafiltration and reverse osmosis processes

Abstract A membrane separation model for tubular module reverse osmosis and ultrafiltration processes was developed in this work. The membrane area of a process can be calculated by this model and the stream matrix of a process can be determined by a process simulation program. In this work the unicorn simulation program was used. The membrane separation model calculates permeate flux and rejection of the solute in small increments of the membrane tube over the entire range of the tube and the process. Calculation of the permeate flux and rejection can be performed by polynomial equations fitted to the experimental data, or by equations based on mass transfer models. The finely porous model and the statistical mechanical model were used. Some experimental data are also needed for determining the parameters of the mass transfer model equations by parameter fitting. The membrane separation model was tested by the simulation of reverse osmosis of aqueous ethanol and acetic acid solutions, and ultrafiltration of aqueous sodium carboxymethylcellulose (CMC) and poly(vinylpyrrolidone) (PVP) solutions. Multistage recycle separation processes with four or five recycle stages were used as test processes. The values of the calculation parameters (i.e., iteration accuracy of the recycle stream, iteration acceleration factor and the number of increments of the membrane tube) giving reliable results with minimum computing time were determined.

Neural Networks for Chemical Engineering Unit Operations

Application of artificial neural networks in simulation of chemical engineering unit operations are studied. On the basis of the present work it is suggested that the raw data describing the phenomenon should be processed so that the relationships of the input variables are studied and if relationships are found, such as dimensionless numbers, these relationships are set as inputs in the neural network. Further, if the phenomenon can be described by a simplified mathematical expression, the value calculated by such simplified solution is set as one of the inputs. The method is demonstrated by three examples, namely, by turbulent fluid flow in a tube, breakthrough curve of adsorption, and by suspension crystallization. In all cases the method presented in this work results in a more accurate simulation and in easier training of the neural network.

The influence of multicomponent diffusion on crystal growth in electrolyte solutions

Abstract Mass transfer phenomena in crystallizing solutes and the concentrations of impurities in solution and in the crystal have been studied based on Maxwell–Stefans theory of multi-component diffusion. The Maxwell–Stefan theory requires data on the driving force, i.e. the concentration gradient for each species. The effect of the electrolyte mixture on the diffusion coefficient of the crystallizing solute and the impurity solute was studied based on the Pitzer theory, which can be used to determine the activity coefficients of a solute in electrolyte mixtures. The studied crystallization systems were the continuous suspension crystallization of potassium sulfate with sodium sulfate as an impurity, and the crystal growth study using potassium nitrate as a single crystal with calcium nitrate as an impurity in the solution. These two crystallization systems have different charge types of electrolytes. The proposed model was applied in calculating the mass transfer coefficient between the crystallizing solute and the impurity solute. Also, the molar flux of the specified species for the studied cases was estimated.