Computer-aided molecular design using genetic algorithms
Abstract
The search for new materials possessing desired physical and biological properties is an important endeavor for designers in the chemical, material and pharmaceutical industries. The traditional approach to molecular design involves a laborious and costly hypothesis, synthesis, and evaluation procedure. To alleviate the protracted design cycle, this thesis proposes a new computer-aided framework for modeling and automating the molecular design process. The proposed approach combines genetic algorithms (GAs) for identifying candidate molecules and group contribution methods for structure-property prediction to evaluate the candidate molecules. Genetic algorithms are general purpose, stochastic, evolutionary search and optimization strategies based on the Darwinian model of natural selection. The essence of a genetic search lies in allowing a dynamically evolving population of molecules to gradually improve by competing for the best performance. In this research, the mechanics, characteristics and viability of using GAs for molecular design is fully elucidated and demonstrated via three industrially important design endeavors: (i) engineering materials design, (ii) semiconductor encapsulation materials design, and (iii) refrigerant design. The results show that the genetic search was able to locate optimal designs for many desired target constraints. It was also able to discover a diverse population of near-optimal designs. The basic GA framework was further extended to better handle constraints such as chemical stability, ease of synthesis, environmental impact, etc. The merits and potential deficiencies of this approach in comparison with other techniques were also discussed. The thesis also investigates the application of neural networks as a possible structure-property prediction module for the GA. Neural networks are dynamic systems composed of highly interconnected layers of simple neuron-like processing elements. The network interconnection strengths are collectively modified which enables the network to create complex nonlinear input-output mappings. Structure-property relationships were developed using the neural network approach by providing input descriptors which reveal aspects of chemical structure that affect macroscopic properties. A comparative study of the predictive abilities of neural networks, group contribution methods, and regression models for density and glass transition temperature properties was performed.
Degree
Ph.D.
Advisors
Venkatasubramanian, Purdue University.
Subject Area
Chemical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.