Work

Identification and modeling of chemical reaction networks

Year: 2022
Type: dissertation
Abstract: Kinetic models provide fundamental information about chemical systems, playing a fundamental role in the development of chemical products, and in the diagnosis and optimization of their respective processes. Despite their central role in Chemical Engineering, these models are sometimes derived based on a postulated set of reactions that supposedly explain observed transformations. One or more mechanistic models are usually suggested. However, the procedures adopted in this task can be based on heuristic aspects and qualitative analysis of the experimental data, producing results that might not be completely satisfactory (and questionable) from the points of view of the structural correctness (reaction network and corresponding kinetic expressions) and the precise/accurate description of the systems analyzed. A systematic methodology is proposed in this thesis for the identification of mathematical models that describe the kinetics of chemical reactions from chemical species concentration data collected in batch experiments. The methodology contemplates 7 steps in which the construction of the model is carried out sequentially, allowing for greater certainty and precision about the identified model. The proposed methodology steps are: (i) experimental data treatment incorporating time-invariant relationships for continuous approximation of discrete data and accurate computation of species fluxes, (ii) data dimension analysis for determining model dimension and thorough understanding of the reaction system topology, (iii) determination of the reaction network superstructure incorporating time-invariant and energetic constraints, and factor analysis for structural model identification, (iv) enumeration of reaction networks using discrete optimization, (v) selection/identification of the network structure using experimental data, (vi) systematic kinetic modeling of the several reaction steps, and (vii) proposal of additional experiment tests based on the model information. Through the incremental model development, the identification of the model structure is performed in an uncoupled manner, first elucidating the structure of the reaction network, and then the structures of the reaction kinetic expressions. Both structure identifications are supported by the differential method, in which linear optimization formulations are used to evaluate the rate of species concentration changes and reaction extents, ensuring the achievement of global optimal solutions. In the first case, from a superstructure of reaction networks, all connections between species are explored, guaranteeing that the best reaction network (composed of an independent set of chemical reactions), describing observed changes in the composition of several species, is found. Regarding chemical reaction kinetics, the identification of the kinetic expressions is performed individually for each reaction step. From a superstructure of reaction kinetic expressions built based on first-principles laws and qualitative analysis of the computed reaction rate profiles, the reaction kinetic expression is identified taking into account the best compromise between fit to experimental data quality and model complexity, according to the Bayesian information criterion Once the structural parts of the model have been identified, the nonlinear final adjustment of model parameters, supported by the integral method, is performed at the end of the methodology using nonlinear optimization, with the advantage of having as initial values for the parameters the ones obtained in the previous step of the methodology (based on the differential method). Thus, in the final phase, the optimal model parameters are obtained according to the maximum likelihood sense, in a simultaneous, bias-free, nonlinear regression procedure. The success of the model structure identification is dependent on good estimates of species fluxes and, consequently, reaction fluxes. For this, a robust data pre-processing method is proposed that incorporates time-invariant relationships as inter-profile constraints, increasing the data accuracy and leading to good fluxes estimates. However, when parts of the model are not satisfactorily identified, additional experimental tests are proposed to elucidate the complete structure of the reaction system under analysis. The development of the proposed methodology is illustrated with the application to four case studies from the literature: the thermal isomerization of alpha-pinene, the catalytic hydrogenation of succinic acid and maleic acid (two separated case studies), and the pharmaceutical case study from Pfizer company. On the basis of the results obtained, it can be concluded that: (i) a systematic model development is required for obtaining models with great confidence that are highly process descriptive and of lower complexity (simpler models) when compared to literature model proposals, (ii) experimental data with high uncertainty may compromise the complete identification of the model structure, requiring additional experiments for allowing a better description of the network structure, (iii) the use of energetic criteria to restrict the network superstructure enables a significant reduction in the number of generated networks, saving time and computational effort in the network synthesis step, (iv) the use of precedence constraints is required for generating consistently connected nonlinear reaction networks, (v) systematic methods for reaction kinetic modeling are essential for accurate model identification, avoiding excessive parameterization and obtaining narrow parameter confidence intervals, (vi) the use of time-invariant relationships in the data reconciliation procedure reduces the noise-to-signal ratio, and consequently increases model identifiability, and (vii) the incorporation of experimental data in the network generation phase to identify plausible structures reduce the number of alternative model candidates to be further analyzed, but at the cost of losing the incremental development of the model. The main contributions of this thesis are: (i) a novel method of data reconciliation, where data approximation is achieved with great accuracy by the incorporation of time-invariant relationships, (ii) mixed-integer linear programming formulations conceived to generate connected linear and nonlinear reaction networks (using precedence constraints), which can be applied to any process synthesis problem described by graphs, (iii) a systematic method for individual reaction kinetic modeling that follows a superstructure-based strategy, enabling the obtainment of parsimonious models, (iv) a methodology for validation of reaction invariants and determination of the data invariant space dimension, and (v) a method for building superstructures of reaction networks considering reaction invariant relationships, energetic criteria, and the checking the consistency of the model with the experimental data. In summary, the proposed methodology is able to identify robust models taking full advantage of experimental data and various optimization techniques. The obtained model increases process knowledge and facilitates process design, scale-up, monitoring, control, and optimization, which, naturally, may be used for improving safety, quality, productivity, and revenues of industrial processes.
Language: English
Open Access status: green
OpenAlex ID: W7047858509
Cites: 184
Cited by: 1
Related to: 9