Integrated Multiscale Chemical Product Design using Property Clustering and Decomposition Techniques in a Reverse Problem Formulation
Type of Degreedissertation
MetadataShow full item record
In recent years, chemical engineers in the Process Systems Engineering (PSE) community have increasingly been using their skill set to solve problems in areas beyond the chemical manufacturing processes, focusing instead on the chemical products themselves. This trend reflects a movement within the field to meet the demands of an increasingly competitive and global consumer products marketplace that stresses being the first to market in conjuction with being the highest volume and lowest cost manufacturer. By definition, chemically formulated products deliver specific attributes to the consumer by manipulating a multitude of separate and often competing mechanisms in molecular architecture that operate over a wide range of length and time scales. Examples of chemical products include performance chemicals, paints, cosmetics, pharmaceuticals, proteins, semi-conductors, and foods, among others. Like process design, computer aided chemical product design is a complex programming problem. Generating, integrating, and managing the information, data, and knowledge at multiple length scales for use in various types of product design problems is a significant undertaking. The traditional approach to managing the complexity of this problem has been to compute information at smaller length scales and pass it to models at larger length scales by removing degrees of freedom (coarse-graining) with the objective being to predict macroscopic properties from molecular information. While often the most accurate method for predicting properties, this simulated approach has two limitations: (1) it has an immense computational cost due to hierarchical nesting and (2) it utilizes a priori knowledge of the molecular architecture (i.e. the number and types of atoms or electrons present). This dissertation covers the development of a novel, alternative approach that allows for the simultaneous design of a chemical product’s molecular architecture across multiple scales using a reverse problem formulation, property clustering, and decomposition techniques. The developed framework is specifically designed to utilize experimental data, parameters, and models since the effectiveness of a chemical product is most often determined by consumer attributes based on consumer preference tests. In this work, three specific methodologies are developed. The first method, Attribute Computer Aided Mixture (Blend) Design (aCAMbD) is an extension of Computer Aided Mixture (Blend) Design (CAMbD) and includes experimental data and regression models, specifically, Scheffe canonical and Cox polynomial models. Necessary adjustments to the original clustering algorithm are identified and the design method is rewritten accordingly. The end result is a method capable of performing a mixture design on any chemical constituent data set across multiple length scales, as long as accurate attribute-component models can be established. A case study mixture design of spun yarn is presented to illustrate the method. The second method developed is Attribute Computer Aided Molecular Design (aCAMD). It is an extension of Computer Aided Molecular Design (CAMD) to include experimental data and regression models, while continuing to use group contribution method (GCM) based property models. The technique uses design of experiments (DOE) to generate an attribute-property relationship and maps the attribute information into a property domain where molecular design can proceed. Adjustments to the property clustering algorithm are made to reflect the new design approach. The result is a method capable of performing molecular design on any attribute data set as long as a strong relationship between attribute and property models can be established. A case study involving the molecular design of environmentally benign refrigerants is presented to illustrate the method. The third method developed is Characterization Based Computer Aided Molecular Design (cCAMD). It is a wholly new method that addresses the limitations of aCAMbD and aCAMD, namely difficulty in finding suitable attribute-component and attribute-property models for complex chemical products. The method uses characterization tools like infrared and near-infrared spectroscopy (IR/NIR) to generate a set of data from a chemical constituent training set and then applies decomposition algorithms like principal component analysis (PCA) to find the underlying latent variable data structure. A parameterization of the data structure into a characterization based group contribution method (cGCM) follows. Attribute data is then mapped into the latent domain using a separate principal component regression (PCR) or partial linear regression on to latent surfaces (PLS) model. A molecular design is then performed in the latent domain. The resulting method is capable of performing structured molecular design across multiple scales for any system of attributes whose molecular architecture can be adequately described by characterization methods. A case study on the particle design of pharmaceutical excipients for an acetaminophen tablet is presented to illustrate the method.