Approaches and Software Tools for Modeling ... - ACS Publications


Approaches and Software Tools for Modeling...

0 downloads 251 Views 4MB Size

Energy Fuels 2010, 24, 58–67 Published on Web 07/21/2009

: DOI:10.1021/ef900488k

Approaches and Software Tools for Modeling Lignin Pyrolysis† Zhen Hou,‡ Craig.A. Bennett,‡ Michael T. Klein,*,‡ and Preetinder S. Virk§ ‡

Department of Chemical and Biochemical Engineering Rutgers University Piscataway, New Jersey 08854, and Department of Chemical Engineering Massachusetts Institute of Technology Cambridge, Massachusetts 02139

§

Received May 19, 2009. Revised Manuscript Received June 29, 2009

The integration of the Attribute Reaction Modeling approach with the CompGen and KME kinetic modeling tools is described through the development of a kinetic model for lignin pyrolysis. Lignin structure is described as the juxtaposition of a methoxy phenol (MP) and a propanoid side chain (PC) attribute at para positions on an aromatic ring. The CompGen tool provides the qualitative list and the quantitative composition of the attributes in the reactant lignin. Pyrolysis alters the state of the MP and PC attributes while leaving the aromatic ring conserved. These reactions are described in two KME submodels, one for each attribute. The juxtaposition of the reaction-altered attributes defines the product slate. This modeling approach describes the full composition of 624 molecules with 50 equations. The userfriendly CompGen and KME tools allow for the construction of the initial conditions and attribute model equations in a spreadsheet environment enhanced with Visual Basic code that creates, compiles and executes the underlying C code.

pyrolysis. Lignin, the relatively refractory component of lignocellulosic biomass, is a phenolic copolymer whose utilization via thermochemical conversion is being studied extensively. In outline of the remainder of this report, we first consider the significance of lignin and lignin pyrolysis. We then use the development of a lignin pyrolysis model both as a contribution toward the utilization of lignin in thermochemical processes and also as a vehicle to demonstrate the more general, feed-independent set of modeling tools. The model development will consider lignin structure, reaction paths and kinetics, pyrolysis network development, and pyrolysis product prediction, in turn.

Introduction The recent spike in the cost of oil has renewed interest in the search for alternate feedstocks for the supply of energy. The utilization of such feeds is complicated by not only the economics of upgrading and conversion but also the environmental footprint, including the impact on CO2 issues. Thus, both primary utilization and derived technical issues need to be addressed. Although each potential resource has its own set of feedspecific technical issues, some common obstacles exist. A significant portion of these alternate feeds are a complex mixture of complex molecules that are often found as covalent or physically aggregated macromolecules. As indicated in the US DOE report on basic research needs for catalysis for energy,1 it is critical to identify the structures and reaction pathways of these heavy feeds and to develop robust computational tools to model their reaction trajectories. This would contribute not only to the processes for utilization of these feeds but also to the design of catalysts to produce efficient and environmentally optimal outcomes. The present report thus has two complementary purposes: the first is to report on the development of computational tools for modeling the structure, reactions, and properties of these feeds and their derived products. In this context, the tools are generic, and their application to lignin, coal, shale, resid, and even gas oils differ only in the quantitative details of the feeds. That is, each of these feeds can be characterized, relative to petroleum naphtha, as carbon-rich, hydrogen-deficient resources laden with heteroatoms and often found in macromolecular form. The second purpose is the specific development of a computational modeling environment for lignin

Lignin as a Resource and a Model Our interests in lignin are both as a possible feed in the biomass contribution to new energy supplies and also as a relatively well understood model of low-rank coals. The energy situation of the 1970s created an intense interest in the use of coal in liquefaction and gasification processes, and as the microbially resistant component of the biomass coal precursor, lignin was examined to shed light on the reaction paths and kinetics involved in coal utilization. Considerably less effort was invested in the utilization of lignin as an energy resource, save the burning of the pulping waste product for energy recovery. These interests are also clear in the current energy scenario, with the exception that ligno-cellulosic biomass is considered a viable option in and of itself. Pyrolysis, in turn, is a fundamental component of the potential thermochemcial conversion processes, such as gasification, pyrolysis, and liquefaction to bio-oils. Thus, the study of lignin pyrolysis modeling seemed cogent. More generally, lignin can be described as a macromolecular feed with aromatic rings, ring substituents, and heteroatoms. Save the details, this also describes coal, resids, and shales, and thus developments in lignin pyrolysis modeling should support the development of models for these other feeds as well.

† Presented at the 2009 Sino-Australian Symposium on Advanced Coal and Biomass Utilisation Technologies. *To whom correspondence should be addressed. E-mail: mtklein@ jove.rutgers.edu. (1) Bell, A. T.; Gates, B. C.; Ray, D., Co-Chairs, Basic Research Needs: Catalysis for Energy, Report from the U. S. Department of Energy. In Basic Energy Sciences Workshop, August 6-8, 2007.

r 2009 American Chemical Society

58

pubs.acs.org/EF

Energy Fuels 2010, 24, 58–67

: DOI:10.1021/ef900488k

Hou et al.

The application of the reaction families to the reactive components of the feed and products can create a network of thousands of reactants and rate laws. In these instances it is often helpful to reduce the parameter set of the model by constraining the kinetic, adsorption, and occasionally thermodynamic constants in the rate laws for each reaction family to follow a correlation based on either the linear free energy or van’t Hoff relationships. For very complex systems, this can reduce the parameter burden from thousands to on order of 10-30. For simpler systems, this may not be necessary and independent individual parameters will be sought. The final component is the property estimation function, which has two key roles. First, it provides the connection between the molecular composition and the set of end-use product properties that motivated the development of the model. It also serves to provide intermediate properties that are used in the construction and solution of the model, such as the enthalpy of formation of some key intermediate species. The advantage of the molecule-based model is clear: molecular structure is the proper starting point for the estimation of component and mixture properties. A model with explicit molecular structures allows the full value of modern physical, organic, computational, and analytical chemistry to be brought to bear on the problem. The downside is that molecule-based models for complex systems are very large. A traditional deterministic model has Nþ2 equations, one material balance for each of the N components along with energy and momentum balances. This downside has been addressed through the development of software tools for the construction, solution, and editing of large models. Some of these will be described below. At a certain level of complexity, however, the number of components N is so large that even automated construction tools will not overcome the hardware and software barriers and execution time demanded. In these instances the essential question is whether the N-component mixture can be modeled with fewer than N material balance equations. An approach toward this end will be developed below as well. To describe these tools and approaches, we turn to the optimally complex feedstock lignin and address steps 1-4 of Figure 1.

Figure 1. Molecular-level modeling steps.

Molecular-Level Modeling Reaction kinetic models serve a traditional engineering purpose and an emerging scientific one. The usual engineering goals of reactor design, optimization, and control are well served by the quantitative information a model provides. Scale up, retro-fitting, and the use of new but like-kind catalysts benefit from such models. An emerging hope is that models can also provide a quantitative, testable understanding of the fundamental chemistry of a process. Here the hope is that the development of new catalysts, solvents, and conversion strategies can be assisted by quantitative models. This requires that the models represent the chemistry faithfully, which requires the use of literal chemical structures and not the lumped components that have served chemical engineers so well for so many decades. The key limitation of lumped models is that the species within the model are without properties, save their defining one. This is, typically, a boiling point or solubility class. Other properties, such a performance, environmental, reactivity, and the like, are obscured by the lack of chemical structure. This has motivated the development of molecule-based models that introduce the true chemical structure of the reacting species. This allows calculation of both end-use properties as well as internal model properties (e.g., transition state properties) for the solution and use of the kinetic model. Molecule-based kinetic models for complex feeds comprise the four main components listed in Figure 1. The first is the structure-composition model for the feed. This provides the initial conditions for the equations that represent the kinetics. For simple feeds this is often a nonissue or a single measurement. For complex feeds, this step usually involves the underdefined problem of transforming a set of measurements into a structural representation and quantitative mole fractions. For heavy feedstocks, the measurements are often global in nature, such as NMR, elemental, and SIMDIS analyses. These global measurements are often supported by more detailed analyses, such as those from various types of mass spectrometry analyses on separated fractions. As developed below, lignins are optimally complex in that their structures are actually very well-known. Once the feed stock has been modeled, the second component creates the reaction model by transforming reactants into products using experimentally discerned reaction pathways. It is convenient that even the most complex feeds, with thousands of reactants and thousands of reactions, will have only on order of 10 different kinds of reactions. These reaction families organize both the reaction pathways and, in favorable circumstances, reaction kinetics.

Modeling Lignin Structure and Composition Native lignin is a phenolic copolymer created through the coupling of enzyme-initiated radicals on the phenolic2 and corresponding resonance positions of the three monomers shown in Figure 2. The resulting copolymer thus contains single-ring aromatic cores bonded covalently by one of a handful of interunit linkages. The classic Freudenberg model3 illustrates the types of these links and further conveys a convenient quantitative distribution of the types. Isolated lignins will deviate from this ideal by virtue of differing initial distributions of the three monomers (e.g., different plant types) and also by structural changes induced by the method of isolation. These changes are well-known in the biochemistry and pulping literature. For the present purposes, the Freudenberg model allows illustration of the CompGen tool for modeling structure (2) Klein, M. T.; Virk, P. S. Modeling of Lignin Thermolysis. Energy Fuels 2008, 22, 2175–2182. (3) Freudenberg, K. Neish, A. C. Constitution and Biosynthesis of Lignin; Springer-Verlag: New York, 1968.

59

Energy Fuels 2010, 24, 58–67

: DOI:10.1021/ef900488k

Hou et al.

Figure 2. Lignin monomer alcohols.

Figure 3. Initial lignin (Freudenberg) structure in terms of MP and PC attributes.

and composition. The monomer alcohols, the cross-linking mechanism, and the lignin pyrolysis mechanism suggest that a natural view of lignin structure is as a set of single-ring aromatics each with two attributes. The first attribute is the type of propanoid side chain (PC) attached to each aromatic ring, and the second attribute is the nature of the phenolic or methoxyphenol (MP) substituent on each aromatic ring. Since pyroylsis leaves the rings intact, a model based on the conservation of rings with reaction-altered attributes seems reasonable. Parsing the Freudenberg structure (or analogous structural representation) provides quantitative values for the statistical distribution of the attributes. The Freudenberg result is shown in Figure 3, for both the PC and MP attribute types. By viewing each unit of the Freudenberg model as the statistical probability of the occurrence of the two attributes, the original Freudenberg model can be represented by the joint probability

of two attributes. Note that both the PC and MP attribute sets have “free” and “etherified” subcategories. The free substituents are bonded to only one aromatic ring, whereas the etherified substituents are part of an interaromatic ring linkage. For other lignins, CompGen’s optimization function allows for the best-fit generation of any quantitative composition. In this manner, an optimal set of attribute probabilities can be generated so that the subsequently generated lignin structure provided a best fit match with the available analytical chemistry. To elaborate, CompGen is an Excel-VBA based program that provides a friendly interface for sampling structural attributes and creating a molecular representation for complex feedstocks. Engaging CompGen will, at first, lead to a user-friendly interactive form, such as that shown in Figure 4, which will allow the user to specify and constrain the attribute types. 60

Energy Fuels 2010, 24, 58–67

: DOI:10.1021/ef900488k

Hou et al.

Figure 4. Attributes sampling in CompGen.

The user form of Figure 4 shows how CompGen allows the modeler to select the number of attribute types (AttributeNumber) and assign each individual attribute (CompName) to an attribute category (CompAttName). Each attribute is also given a unique identifier (CompID). In the example of Figure 4, the five attribute types are propanoid chain (PC), methoxyphenol (MP), and the free gas molecules CO2, CH4, and CO. CompGen can represent attributes as independent values or constrain values to follow a probability distribution function (PDF) with limited parameters. Either way, it will optimize a χ2 objective function with chemical analysis terms PEij illustrated below to determine the optimal set of attribute probabilities of PDF parameters for both MP and PC sets. !2 X X PEij -PM ij 2 χ ¼ σ ij j i

attributes mole fractions are to be represented by a probability distribution function. The thus-optimized attribute values will be the final representation of the starting lignin and used as the initial conditions in the kinetic modeling. Lignin Reaction Pathways and Kinetics The structural analysis provides identities of the initially reactive moieties during lignin pyrolysis. Klein and Virk1 studied the set of reactants listed in Table 1 and ascertained reaction pathways, kinetics, and mechanisms. The pathways of the initial reactants suggested additional species to reveal secondary lignin reaction paths, and so on. The converged database of Table 1 thus provides the basis for the subsequent kinetic model: the pathways reveal the reaction network and the rate parameters the kinetics. The reactants listed in the first column of Table 1 mimic the PC and MP attributes as well as the interaromatic unit links (IL) whose fragmentation leads to molecular weight reduction and the formation of single-ring phenolic products. Among the IL models is phenethylphenylether (PPE), a betaether mimic whose main reaction path, illustrated in column 5 of Table 1, is to the hydrogen-balanced product pair phenol and styrene. Table 1 also shows that styrene undergoes secondary reaction to toluene, ethylbenzene, benzene, and higher molecular weight adducts. Subsequent studies of VGE,4 a beta ether with the PPE backbone and additional lignin-like substituents shown in Figure 6, revealed its major

Consequently, CompGen will provide the optimal set of attributes as an input data list for further reaction modeling. The computational flow of CompGen is illustrated in Figure 5. At first, CompGen will provide a form and allow the user to select the MP and PC attribute types of the lignin to be modeled. If the quantitative mole fractions can be determined directly, CompGen will launch a straightforward mathematical parser to convert the molecular compositions to the values of each attribute types. For cases where the molecular compositions cannot be obtained directly, CompGen will invoke the optimization loop of Figure 5. In this case, the objective function will be optimized by adjusting the independent mole fractions or a set of limited parameters if the

(4) McDermott; John, B.; Klein, M. T.; Obst, J. R. Chemical Modeling in the Deduction of Process Concepts: A Proposed Novel Process for Lignin Liquefaction. Ind. Eng. Chem. Proc. Des. Dev. 1986, 25 (4), 885– 889.

61

Energy Fuels 2010, 24, 58–67

: DOI:10.1021/ef900488k

Hou et al.

Figure 5. Flow sheet of CompGen strategy.

reaction path to be dehydration and formation of the PPE analogue with a double bond in the linkage. The fragmentation of this PPE analogue would not be in hydrogen balance. Likewise, the cracking reactions of the alpha ether, phenyl ether, diphenyl methane, and diphenyl are not in hydrogen balance, so their fragmentation to light products consumes hydrogen elsewhere and is thus generally accompanied by the formation of an adduct. The key MP mimic is guaiacol, which follows two reaction paths, one to methane and catechol and the other to phenol plus CO and H2. The other MP models, namely, 2,6-dimethoxyphenol, veratrole, and anisole, follow formally similar routes. The key PC mimics in Table 1 reveal routes to light gases (CO, CO2, H2), water, and various smaller hydrocarbon substituents. In sum, Table 1 suggests that lignin pyrolysis products should include light gases (CH4, CO, CO2, H2), light liquids (H2O, MeOH), a collection of single-ring phenolics (variously substituted phenols, guaiacols, and syringols), and a hydro-

gen-deficient char. The Arrhenius parameters in columns 6 and 7 of Table 1, when incorporated in the material balance equations for each component, provide an a priori basis for the prediction of the kinetics of the formation of these key products. It is thus relevant to now turn to the issue of the lignin pyrolysis reaction network. Attribute-Based Reaction Modeling The reaction products just noted arise as the result of the changes in state of the PC and MP substituents on the conserved aromatic ring. The number of implied molecules for even this relatively well contained two-dimensional (e.g., two attributes) problem exceeds 624. For other feeds and more complex lignins this number will surely grow larger. Thus it seems worthwhile to develop methods to model the dynamics of the molecular composition with fewer than one equation per molecule. For that purpose we turn to the reactions of the attributes. 62

Energy Fuels 2010, 24, 58–67

: DOI:10.1021/ef900488k

Hou et al.

Table 1. Model Compounds for Lignin Pyrolysis

Figure 7 illustrates the attribute reaction modeling approach developed in support of this goal. As shown in Figure 7, a molecular composition with N attribute types with M attribute values will have MN molecules. The associated deterministic reaction model will thus have MN balance equations. This can exceed 50 000 for petroleum resids and

like feed stocks, which is too large for many current applications. The essential idea of the ARM approach5,7 is to treat the reactions of the attributes as independent. The ARM is thus a hybrid model of N submodels, each submodel (6) Wei, W.; Bennett, C. A; Tanaka, R.; Hou, G.; Klein, M. T. Detailed Kinetic Models for Catalytic Reforming. Fuel Process. Technol. 2008, 89, 344–349. (7) Campbell, D. M.; Bennett, C. A.; Hou, Z; Klein, M. T. AttributeBased Modeling of Resid Structure and Reaction. Ind. Eng. Chem. Res. 2009, 48, 1683–1693.

(5) Hou, Z.; Bennett, C. A.; Klein, M. T. Attribute Reaction Model (ARM) Approach for Heavy Hydrocarbon Reaction Modeling; 235th ACS Meeting, New Orleans, 2007.

63

Energy Fuels 2010, 24, 58–67

: DOI:10.1021/ef900488k

Hou et al.

having the number of equations equal to the number of values the attribute can take. The overall problem, then, scales as N  M instead of MN, which is a considerable savings of CPU time. This is the approach followed in the lignin pyrolysis model, which allows convenient illustration since N=2. For lignin pyrolysis modeling, two independent sets of ODEs are used to describe the variations of MP and PC attributes. The temporal variations of all MP substituents can be represented by M equations as: Attribute 1 Submodel:

of attributes. Sampling these updated PDFs provides molecular information for the products and the juxtaposition of all combinations of attributes.

dMP1 dMP2 dMPM ¼ R11 ¼ R12 ::: dt dt dt ¼ R1M

Likewise, the variations of all PC substituents (PC) can be represented by N equations as: Attribute 2 Submodel:

dPC1 dPC2 dPCN ¼ R21 ¼ R22 ::: dt dt dt ¼ R2N

Given initial conditions (e.g., from CompGen), these equations can be solved to obtain a reaction-altered distribution

Figure 8. Flow sheet of KME strategy.

Figure 6. VGE pyrolysis pathways.

Figure 7. PDF sampling for molecular compositions representation.

64

Energy Fuels 2010, 24, 58–67

: DOI:10.1021/ef900488k

Hou et al.

Figure 9. KME model definition options.

Figure 10. Two ARM reaction sets in KME.

interface6 that transforms a list of reactions into solvable code in a format that allows for seven execution modes (once

This logic has been incorporated into Kinetic Modeling Editor (KME), a modeling tool with an Excel-based graphical 65

Energy Fuels 2010, 24, 58–67

: DOI:10.1021/ef900488k

Hou et al.

Figure 11. Pyrolysis products yield profile along PFR.

the creation of N independent kinetic models with shared information, such as reaction conditions, inlet flows, and the like, and a final product parser that creates the molecular composition as the juxtaposition of attributes. This is illustrated in Figure 10 for the lignin case at hand, which depicts two KME reaction models, one for the reactions of MP and one for the reactions of PC attributes. The current ARM model of lignin pyrolysis thus comprises the following steps. First, KME obtains the initial MP and PC attribute values from CompGen. It then generates the ODE values for the MP and PC submodels. The user then enters the input and output data, along with the reaction conditions and rate parameters, which allow the model to be run. The reaction-altered attributes are then juxtaposed to create the final product slate of light gases, light liquids, phenolics, and char.

through in steady state, parameter estimation in steady state, goal seeking in steady state, once through in deactivation mode, parameter estimation in deactivation mode, goal seeking in deactivation mode, and goal seeking with parameter estimation in deactivation mode). KME can use a simple reaction notation (e.g., A þ B f C) to create the ODEs and computer code necessary for running the model. In addition, KME provides a flexible modeling environment that provides for various reactor types and flexible rate-law editing. The computational flow of KME is illustrated in Figure 8, whereas Figure 9 illustrates the key early options the KME user selects. Engaging KME opens a set of dialogue forms that allow the user to set up modeling options, such as the execution mode, reactor type, and energy balance details. In the “REACTION” sheet, users can either enter the reactions following the KME syntax or import a reaction list from other software, such as NetGen/InGen. KME will build the model and generate its C code automatically. After the model is built, the user can enter the model data: feed data in the “INP” sheet, conditions in the “COND” sheet, and product measurements in the “OBS” sheet. The user can then launch the model to get the numerical solutions. After the model is solved, KME provides a set of tools to analyze the results. This includes species’ profiles through the reactor, a parity plot, statistical analysis, and so on. The foregoing describes the basic one-dimensional KME. Essentially, its extension to include the ARM strategy involves

Results: Application to Lignin Pyrolysis The lignin pyrolysis ARM model was constructed using the Freudenberg model for attribute input, the database of Table 1 for reaction pathways and kinetics, and the KME input sheets illustrated in Figure 10 and supplied separately as Supporting Information. The reaction products, that is, the juxtaposed attributes, are organized into four fractions: (1) light gases, (2) light liquids, (3) single-ring phenolics, and (4) char, which is defined by difference. Representative results, intended only to show 66

Energy Fuels 2010, 24, 58–67

: DOI:10.1021/ef900488k

Hou et al.

Figure 12. Asymptotic yields of key products parametric in the coniferyl/sinapyl alcohol ratio in starting lignin.

the nature of the CompGen/KME outputs, are shown in Figure 11 and 12. Figure 11 depicts the reactor profile of the yields of the key products CH4, CO, phenol, guaiacol, and syringol. Figure 11, depicting the asymptotic yields of CH4, CO, phenol, guaiacol, and syringol as a function of the starting lignin’s coniferyl/sinapyl alcohol ratio, shows the integrated nature of the tools and the link between CompGen and KME. Model predictions can be compared to both feed characterization and reaction experiments. The model parameters will generally be tuned to provide an optimal fit. It is standard to adjust the attribute probabilities in the initial feed to optimize the match between the model and feed characterization, whereas, for instances where the combined feed-reaction model is being compared to reaction results, the ordinary practice is to adjust the Arrhenius parameters.

Conclusions An integrated tool set for the modeling of complex feeds was illustrated via its application to lignin pyrolysis modeling. The statistical approach embodied in ARM paradigm provides a feasible solution for maintaining molecular level detail without the burden of one ODE for every molecular species in the model. By integrating the ARM approach with the user-friendly KME and CompGen tools, the complex lignin pyrolysis model can be created, solved and edited quite easily. This allows for engineering simulation of a wide range of lignins as well as for testing “what if” scenarios for lignin structure or manipulation strategies. Supporting Information Available: This material is available free of charge via the Internet at http://pubs.acs.org.

67