Multivariate Characterization, Modeling, and Design of Chemical Products in Property Cluster Space

Abstract

The focus of this dissertation is on the development of in silico approaches for the logical and systematic solution of chemical product design problems. The application of multivariate characterization, modeling, and design is accomplished by utilizing interdisciplinary methods and tools that extend through multivariate statistics, applied mathematics and computer science. Methodologies and techniques such as spectroscopy-based group contribution methods, chemometric/chemoinformatic techniques, reverse problem formulation, and property clustering techniques are integrated within computer-aided molecular/mixture design (CAMD) algorithms to design chemical products in a computationally efficient manner that provides optimum performance in terms of customer requirements. Property-based design techniques and multivariate data-driven modeling and optimization strategies are presented in this dissertation covering two specific areas of chemical product design: mixture and molecular design. In mixture design, the property integration framework is combined with multivariate statistical techniques and applied in a reverse problem formulation on chemical product design problems by systematic and insightful use of past data describing the properties of the raw materials, their blend ratios, and the process conditions during the production of a range of product grades to achieve new and improved products. Projection methods, like principal component analysis (PCA) and partial least squares (PLS) are applied to identify the underlying relationships necessary for simultaneous optimization of all three variables. The method is illustrated using a polymer blending problem. iii In molecular design, multivariate characterization techniques like infrared (IR) spectroscopy are utilized to generate numerical descriptors of molecular architecture in terms of IR frequency of a set of representative samples. Models based on quantitative structure-property relations (QSPR) are used to elucidate structure-property relationships. Applying principal component analysis, high dimensional and highly correlated molecular descriptor variables are transformed into low dimensional and statistically independent latent variables. These latent variables are then used to calibrate latent property models. Finally, the reverse design of molecules is accomplished by exhaustively searching for molecular structures with target properties, from the combinatorial building blocks. A characterization-based group contribution method is utilized to estimate the properties of the formulated chemical products. The concepts and the solution methodologies are demonstrated using two proof-of-concept examples: biodiesel additive formulation and ionic liquid design.