Efficient approach to optimization under uncertainty with application to large scale engineering systems.

摘要

Increasing model complexities of various engineering systems, coupled with a growing number of design choices creates a challenging problem of finding the optimal design of the engineering system with minimum computational efforts. Additionally, real-life large-scale engineering systems are prone to several sources of uncertainties, like model prediction imprecision. These further hinder the process of obtaining a reliable and robust design. Furthermore, huge CPU requirements of the system model restrict conventional optimization methods to comparatively smaller scale engineering problems, invariably ignoring uncertainties. This research focuses on the development of an efficient optimization framework under uncertainty, for engineering system models of varied complexities.; The main contributions of this work include: (a) First successful application of genetic algorithms for optimal design of heat exchangers, using black box models; (b) A parallel optimization framework using Supercomputers for computationally efficient optimal heat exchanger design; (c) A novel sampling approach to stochastic optimization for Computer Aided Molecular Design (CAMD) under property prediction uncertainty, enabling sensitivity analysis of the model parameters and obtaining optimal molecular designs for desired properties; (d) Improved computational efficiency of the Monte Carlo algorithm for molecular simulations by using a new highly-efficient sampling technique replacing costly Monte Carlo samples; (e) Improved overall efficiency and accuracy of the generalized engineering systems and optimization framework for optimization under uncertainty at the levels of the model, sampler, and the optimizer.; The modeling environments considered include the black-box model for heat exchanger design from the Heat Transfer Research Institute (HTRI), simpler and transparent group contribution models (GCM) for property prediction of polymers, complex molecular simulation models based on statistical mechanics for property prediction of molecules and the Computer Aided Engineering (CAE) models used in manufacturing industries, that also act as a black box.