NEURAL NETWORK MODELING OF TRANSITION METAL - ZEOLITE EXHAUST CATALYSTS

Recent research in automobile exhaust catalysts has addressed the substitution of platinum-group metals Pt, Pd and Rh by metals such as Cu, Co, Ag, Zn, Mn and Sr exchanged or impregnated on zeolites, TiO₂ or Zro₂ carriers. These catalysts have the potential of becoming good alternatives to the commercial three-way catalysts to convert pollutant hydrocarbons (HC), carbon monoxide (CO) and nitrogen oxides (NO_x). This paper describes a technique based on neural networks, to correlate the catalyst synthesis variables and resulting exhaust conversion. The optimum catalyst composition and operating conditions for a specified exhaust conversion are determined

A back-propagation algorithm was used to train the feed-forward neural network consisting of two hidden layers with 45 and 60 neurons in the first and second hidden layers respectively, with optimum values of the learning factor and momentum gain coefficient. The effects of the operating and compositional parameters on NO_x conversion by Cu-ZSM-5 were found. The optimum conversion was predicted for Si/Al atom ratio in the range 30-35, Cu-loading (in Cu-ZSM-5) of 1·1 - 1.2% of the zeolite weight, and an operating temperature of 650-675  K. The rare-earth metals (Ce, Cs and La) that act as promoters for three-way catalysts did not have a considerable effect on the exhaust conversion. The conversion increased by at least 10% when Co is used as a co-cation in Cu-ZSM-5.     

A NONLINEAR OBSERVER BASED ON HYBRID MODELLING OF CHEMICAL REACTORS

The successful design of an observer for inferring the outlet composition from a chemical reactor heavily relies on the goodness of the adopted kinetic rate model (Baratti et al., 1993). On the other hand, often, it is difficult to dispose of a simple, but, exhaustive kinetic model because of the complexity of the reaction scheme one has to deal with.

In this work, we explore the possibility to represent global (lumped) reaction rate laws by the use of neural network models. The aim is to develop a nonlinear observer (extended Kalman filter, EKF) of an heterogeneous gas-solid reactor that relies on a grey model where the “neural reaction rate” law is integrated within a first principles model. The procedure is outlined for the case of the catalytic oxidation of carbon monoxide over Pt-alumina catalyst. The results show that neural networks (NN) can be effectively used in representing lumped reaction rates since NN are able to capture the essential characteristics of the functional relationship relating the state variables.     

Heterogeneous Catalytic Oxidation of Acrolein to Acrylic Acid: Mechanism and Catalysts

Partial oxidation of acrolein is a commercially important reaction, its product—acrylic acid—being widely used industrially for producing resins, dyes, glues, nonwoven fabrics, etc.

Partial oxidation of acrolein is also a convenient model reaction because: (1) the number of reaction products is moderate (CO, CO₂, acrylic acid) and (2) their difference in acid-base properties from the starting material makes it possible to select desirable catalysts by applying directly and efficiently Boreskov's concept of intermediate chemical interaction of a catalyst with reaction mixture components. According to this concept [1], the transformation of surface intermediates (SI) formed in the interaction of reactants with a catalyst's surface is determined by the structure and bond energy of these SI.

The study of the reaction mechanism includes determination of structures and energy characteristics of the surface intermediates and the elucidation of their connection with catalyst chemical composition and reaction routes to particular products. This reliable information helps us to understand the nature of catalyst action and to elaborate the theory of catalyst selection. We have used this method to approach the problem of the systematic selection of catalysts for the oxidation of acrolein to acrylic acid. The review summarizes the research done in the lnstitute of Catalysis of the Siberian Branch of the Russian Academy of Sciences during recent years.     

OPTIMAL CATALYST DISTRIBUTION AND DILUTION IN NONISOTHERMAL PACKED BED REACTORS

The problem of the optimal distribution of two different catalysts in a nonisothermal packed bed reactor is analyzed. State variable constraints, such as maximum allowable temperature, may be incorporated through a penalty function approach into the steepest ascent solution of this problem.

Loading a wall-cooled reactor with two chemically different catalysts is compared to dilution of a single catalyst with inerts for temperature moderation. A criterion is presented for estimating whether catalyst dilution may be necessary in order to keep reactor temperature below some maximum allowable value.

Conditions for optimal loading with a two-dimensional reactor model are presented and applied to catalyst dilution in a butane oxidation reactor. Suboptimal catalyst loadings are compared with the optimal loading.     

Catalyst design using artificial intelligence: SO2 to SO3 case study

Catalyst design is key to the improvement of chemical process efficiency. The required work for the development of new catalysts can be supported through the proper application of artificial intelligence to identify optimal compositions. A generic methodology for the application of machine learning to catalysis research is therefore outlined in this work. The catalytic oxidation of SO₂ was used to exemplify the first iteration of this methodology. 1784 data points from 31 published papers were compiled into a databank. The inlet SO₂ concentration ranged from 0 to 66 mol%. An artificial neural network (ANN) was trained on the databank in order to predict SO₂ conversion based on the catalyst composition and the reactor operating conditions (temperature, pressure, catalyst mass: volumetric flowrate ratio (w/v), and feed composition). The model achieved a root-mean-square error of 6.6%. A preliminary screening step identified 3:1 V-Mg/SiO₂ catalysts as exhibiting high conversion at 648 K. A multi-objective optimization was then performed on a single catalyst to identify solutions exhibiting high conversion and high productivity at 648 K while minimizing the catalyst cost. The optimal solution was predicted to be a 2.9 wt% V/0.2 wt% Mg/SiO₂ catalyst operating at a w/v of 7.49 kg-cat · s/m³ STP, achieving 100% SO₂ conversion with a material cost among the bottom third of cost values. Artificial intelligence can then be employed to extract useful knowledge from published catalytic data and orient future search for novel catalyst development.     

MULTIPLE CATALYSTS PRODUCE A SYNTHETIC FUEL FROM COAL

A recent government study indicates that oil prices will be substantially higher within the next few years. Thus liquid fuels from coal, will be required. The study concludes, however, that the considerable effort, time and money spent to date in trying to make gasoline from coal will be of no help in reducing in the 1990's our almost entire dependence on foreign oil.

Any grade of coal may be gasified with oxygen. The syngas is separated and adjusted, to give substantially a mixture of hydrogen, carbon monoxide, carbon dioxide after sulfur removal, and adjustment of the carbon to hydrogen ratio for methanol. It is then passed at an elevated temperature and pressure through a catalyst bed. A small percentage is converted to methanol with the evolution of a large amount of heat. This heat is removed to prevent overheating the catalyst and, in the past, has been wasted to a large extent. The gas has been recycled to build up a maximum concentration of 3 to 4% methanol containing on condensation 10 to 20% water, which must be separated.

For the American (Wentworth) process, a second catalyst has been developed whose optimum reaction temperature is above that of the gas stream leaving the first catalyst. The gas stream is passed through this second bed of catalyst without cooling, and on through a third bed of previously developed catalyst, whose optimum reacting temperature is still higher. The accumulated heat of the gas stream leaving this third catalyst bed is at a high enough temperature to allow the recovery, along with waste heat from the gasifier steps of most of the reaction heat in a waste heat boiler together with that from subsequent passage of the gas through additional beds of the first two catalysts. This recovered heat supplies energy for the complete operation of the plant including coal handling and oxygen preparation, as well as, for some electric power for sale outside.

Flow sheets with temperatures and material and heat balances show the reasons of the advantages of the multiple catalyst process, including the considerable saving of energy accomplished in the much lower compression needs for recycle of the much lower amount of gas through the beds of catalysts.     

DEVELOPMENT OF RANEY-TYPE LOW TEMPERATURE METHANOL SYNTHESIS CATALYSTS†

Catalysts prepared by caustic leaching of Cu—Zn—Al alloys are shown to have activities greater than those of a commercial copper-based catalyst.

Complete leaching of alloys containing 50 wt.% aluminium, with between 0 and 50 wt.% copper, and the balance zinc produced catalysts having a wide range of activities. The most active catalysts for methanol production were produced from alloys containing from 10 to 20 wt.% zinc.

The activities of catalysts prepared by the partial leaching of an alloy containing approximately 36% copper, 15% zinc and 48% aluminium were shown to pass through a maximum at a leaching time of 2.75 hours. This catalyst had more than double the activity of a commercial catalyst al reaction conditions similar to those employed industrially.     

PERFORMANCE OF A PACKED-BED REACTOR

We examined the conversion rates in a packed bed catalytic reactor with a two phase upward flow in a wide range of operating conditions. The oxidation of ethanol to acetic acid in the liquid phase on a Pd-Alumina catalyst was chosen as the test reaction.

Global reaction rates were measured by changing gas velocities, temperature, and feed concentrations of ethanol in the liquid phase. The observed rates were compared with those calculated using two models, assuming a total external wetting of the catalyst. In the first model, a “kinetic” conversion rate was calculated by neglecting any interphase mass transfer resistance. In the second model the interphase mass transfer resistance was considered and expressed by an overall coefficient evaluated from published correlations. The results show that there is an hydrodynamic influence, probably due to the mass transfer and/or to the partial effective wetting of the catalyst. Mass transfer, on the other hand, is better than that observed in other cases. A comparison with the performances of a downflow trickle-bed reactor operating at the same tested conditions showed a much smaller influence of mass transfer and hydrodynamics on the overall conversion rate for the upflow reactor.     

Modeling preferential CO oxidation over promoted Au/Al2O3 catalysts using decision trees and modular neural networks

In this work, the experimental data for CO oxidation over promoted Au/Al₂O₃ catalysts were analyzed using decision trees and modular neural networks. The full dataset was first classified by decision trees to identify and select the conditions for high catalytic activity; then the reduced dataset containing only the promising data were modeled using neural networks, at which the compositional and operating variables were processed in a modular manner to be able to model their effects together but treat them separately. It was found that operating variables were more influential on catalytic activity than catalyst compositional variables. The temperature was found to be the most significant operating variable while Mg and Mn were the best performing promoters. It was also concluded that decision trees and neural networks can complement each other to extract easily comprehensible knowledge, and they can be used for similar catalytic systems for the same purpose.     

Deactivation regeneration and poison - resistant catalysts: commercial experience in stationary pollution abatement

The design of a commercial catalytic system requires considerable information on the operating conditions and environment in which the catalyst will function. Feed concentrations, flow rates, temperatures, pressures, and other measurements are necessary to establish a preliminary design for the required steady-state conversions. In addition, the catalyst and process design must be capable of functioning after experiencing upsets in any of the above mentioned variables. The most common causes of catalyst failure are thermal deactivation and poisoning by constituents in the gas stream.

This paper describes factors leading to deactivation of catalysts in commercial stationary abatement installations. Deactivated catalysts are returned and laboratory characterization methods are used to identify failure modes. These results are used to modify the catalyst, develop methods for regeneration and to recommend plant operating conditions.     

THE SYNTHESIS AND EVOLUTION OF NETWORKS OF HEAT EXCHANGE THAT FEATURE THE MINIMUM NUMBER OF UNITS

An important problem in the design of chemical processes is that of bringing process streams from their temperatures of availability to the temperatures at which they are needed without undue cost. An important strategy for reducing the cost of doing this is heat recovery: Using the heat available from streams to be cooled to service streams to be heated.

In the absence of nonthermodynamic constraints, it is not difficult to assess the amount of heal recovery possible; and methods have been proposed (Linnhoff and Flower, 1978) that allow full heat recovery to be systematically obtained. The networks to which these methods lead are, however, more complex than necessary. Typically, therefore, those methods have been augmented with techniques for the evolutionary development of the initial network in order to simplify its structure, usually by minimizing the number of “matches” between streams.

The present work proposes a simple method for exploiting features of maximally simple networks (those that have as few “matches” as possible) in order to design, with greatly reduced effort, such a network that features a high (typically complete) degree of heat recovery. Further exploitation of those features allows a simple method of evolutionary development that makes it possible, in a very few evolutions, to improve the initial network as much as the data allow.

Unlike the other methods offered, the present ones are not hindered by the presence of nonthermodynamic constraints (practicality, safety, operability). Their generality and enhanced simplicity make them, more than any others, applicable in an industrial context. Although intended for application by hand, they lend themselves admirably to computer implementation, especially in an interactive mode.

These methods are demonstrated on the standard test examples and prove themselves powerful.     

INFERENTIAL SENSORS FOR ON-LINE MONITORING OF TISSUE MACHINE QUALITY PROPERTIES

There are many processes in a pulp and paper mill where an on-line parameter analyzer cannot be used due to several reasons

•The analyzer is very expensive

•It cannot survive in the environment we want to use it

•It is not operational due to hardware problems, maintenance etc

•Such an analyzer does not exist in the market

In all these situations it would be great for the mill to have an alternative way of measuring those parameters in real-time. Neural network models can serve as virtual sensors that infer process parameters from other variables, which can be measured on-line. One excellent application of inferential sensors in the pulp and paper industry is the on-line prediction of paper properties, like tensile, stretch, brightness, opacity etc. In tissue machines, the most important quality parameter is softness, which is usually measured in a very subjective manner by the touch of a human finger. In this work we examine how neural networks can be deployed in order to build online virtual sensors for softness and other tissue quality properties. The results are promising and show that neural network technology can improve productivity and minimize out of specs production in a tissue machine, by providing accurate real time monitoring ofquality parameters.     

EXTENSION OF THE NEURAL NETWORKS OPERATING RANGE BY THE APPLICATION OF DIMENSIONLESS NUMBERS IN PREDICTION OF HEAT TRANSFER COEFFICIENTS

The paper presents a study aimed at extending the neural network mapping ability. In traditional modelling, operational process parameters (gas/material temperature, air velocity, etc.) are the inputs and outputs to and from the network. In this approach dimensionless numbers (Re, Ar, H/d) were used as inputs to predict the heat transfer coefficient in a fluidised bed drying process. To produce the data set necessary to train the networks, drying trials of different materials in a fluidised bed were carried out.

A series of simulations were performed and several neural networks structures were tested to find an optimal topology of the network. Training data set contained information only about two materials. The networks were tested using data obtained for the third product.

Performance of the network was satisfactory, however further improvement of mapping ability may be expected after filtration of the testing data.     

CONTROL OF A PULSED LIQUID-LIQUID EXTRACTION COLUMN BASED ON A MULTILEVEL SYSTEM OF AUTOMATA

This paper presents the real-time application of the learning control theory to the control of a chemical pilot plant: a pulsed liquid-liquid extraction column.

The behaviour of an agitated liquid-liquid extraction column can be related to random mechanisms such as the phenomena of droplets breakage and coalescence. Previous studies on hydrodynamic and mass transfer aspects showed that a pulsed liquid-liquid extraction column had an optimal behaviour for operating conditions close to flooding. These results led to choose the following strategy to control the column in its optimal behaviour zone:

- the measure of the conductivity of the liquid medium below the distributor which gives a good information about flooding, is the controlled variable

-the pulse frequency is the control action.

The learning control algorithm is based on a multilevel system of automata which operates in a random environment. By means of an evaluation unit of the performances of the column which generates either penalty (inaction) or reward on the basis of heuristic rules, the automaton chooses a value of the pulse frequency. This approach is essentially connected to artificial intelligence in so far as human knowledge on the plant is included in these rules.

This algorithm has been implemented on a microcomputer for control purposes. The experimental results presented show the good performances of the approach.     

OPTIMIZATION OF THE OPERATION IN A REACTOR WITH CONTINUOUS CATALYST CIRCULATION IN THE GASEOUS BENZYL ALCOHOL POLYMERIZATION

Polymerization of gaseous benzyl alcohol was studied on a commercial zeolite (MZ-7P supplied by Akzo-Chemie) in a jet spouted bed reactor which is an original gas-solid contact technique whose use in this type of processes has not yet been described in relevant literature.

Properties of the catalyst-particles-bed and polymer production are calculated by a simulation routine as a function of the following operating variables: the partial pressure of the benzyl alcohol fed into the reactor and the average residence time of the catalyst particles in the reactor. The simulation method and the kinetic equations used have been checked by experimental data.

An analysis of the effect or the operating conditions is carried out, thus determining the values which enable maximum production.     

EFFECTS OF SELECTED CATALYSTS ON KINETIC PARAMETERS OF THE CARBON-STEAM REACTION

A comparison has been made between the catalytic effects of two groups of elements on the steam gasification of carbon: the alkaline earths and noble transition metals of Group VIII. Spectroscopically pure graphite was selected as a model carbon because it consists of well-defined crystals with a large number of atoms located in basal planes, which, without a catalyst, are of low reactivity. Moreover, foreign atoms which can enhance or inhibit catalysis are eliminated. Local overheating is prevented by the endothermic reaction with steam.

It is not possible to distinguish unambiguously reaction-rate increases associated with a lowering of the activation energy and those caused by an increase of the reaction-site density. Thus conclusions about the mode of catalyst action are by necessity tentative. However, evidence from electron micrographs, in connection with elemental maps and theoretical calculations indicates that catalysis in the case of alkaline earths can be explained essentially by an increase of the reaction-site density. A close contact between these catalysts and the carbon surface is preserved during the reaction. In contrast, metals of Group VIII have a tendency to coalesce and to become isolated from the graphite surface. Here a lowering of the activation energy is a distinct possibility due to intercalation of these metals and weakening of the C-C bonds which are to be severed in the carbon gasification.     

CARBON COMBUSTION RATES AND TEMPERATURES IN SHALLOW FLUIDIZED BEDS

The mechanism of combustion of carbon in shallow fluidized beds at temperatures 750-1000°C is studied by measuring burning rates and temperatures of spherical carbon particles ranging from 2 mm to 12 mm diameter directly in an experimental fluidized bed. Among variables investigated were inert particle size, superficial fluidizing velocity, temperature, the influence of neighbouring active particles and oxygen concentration in the fluidizing gas.

Under the experimental conditions explored, combustion was mainly kinetically controlled, so that with carbon particles larger than about 4 mm, burning rates are significantly higher than those predicted by combustion models which assume combustion to be controlled by the rate at which oxygen diffuses through a stagnant particulate phase surrounding the burning particle. The higher burning rate seems to arise because the greater mobility of particles in the bed causes the restriction to oxygen flow to the carbon surface offered by the particulate phase to be reduced and has important consequences for combustor design.

Measured carbon particle temperatures were influenced considerably by bed operating conditions ranging from 15 to 215°C higher than bed temperature.

Measured burning rates of carbon particles were found to be reduced significantly when other active particles were present in the bed. This sensitivity of burning rate to changes in active particle concentration in the bed was shown to be increasingly important once the concentration of carbon in the bed exceeded about 1%

Increasing the bed inert particle size, superficial fluidizing velocity, oxygen concentration in the fluidizing gas and bed temperature resulted in higher burning rates. The implication of these findings on combustor design are discussed.     

OXIDATION OF SULFUR DIOXIDE ON WATER-REPELLENT ACTIVATED CARBON

To reduce resistances of interphase and intraparticle mass transfer in an aqueous solution, water-repellent catalysts were prepared by spraying polytetrafluoroethylene (abbreviated as PTFE) on activated carbon.

Oxidation of sulfur dioxide on activated carbon in the presence of water was chosen to elucidate the effect of water-repellent catalysts on the performances of gas-liquid cocurrent upflow and downflow reactors. The reaction rate increases and reaches a maximum with the increase in PTFE loading. The enhancement of reaction rate by PTFE may be explained by the decrease of partial wetting on the surface     

ABSORPTION AND DESORPTION TO AND FROM LIQUID FILMS ON INCLINED PLANES

Two published theoretical models are examined and applied to experimental results for absorption and desorption. The system used was CO₂/H₂O and studies were made for liquid film flow down inclined planes. Experimental results give “Reduced” values of mass ransfer rates.

Interferometric studies give interfacial concentration, penetration and film depths, and take-up of carbon dioxide. In the case of desorption the interferograms are distorted by “deflections.”

All the experimental values for absorption and desorption differ from those calculated from theoretical models.

Desorption is not a mirror image of absorption, and it is approximately 75% of the transfer rate of absorption over a wide operating range.

A comparison is made of the behaviour of static pools and flowing liquid films.     

UCKRON II TEST PROBLEM THE EFFECTS OF PORE DIFFUSION

The UCKRON I test problem was reworked with the inclusion of diffusional limitations to the kinetics. For this purpose, the rate expressions, i.e. the full kinetic model and the Langmuir-Hinshelwood model, proposed in the previous workshop, were assumed to represent the true kinetics.

A simple mathematical model based on simplified effective diffusivities and the “Dusty-Gas” model were used to take account of pore diffusion resistance within the catalyst particle. These models predicted effectiveness factors based on the reaction at the bulk conditions. The effectiveness factors were then correlated against temperature and the partial pressure of reacting species.

These correlations were then used to predict the temperature profiles within the water cooled plug flow reactor for various values of cooling media temperature.