Efficient Global Optimization Techniques for Thermohydrodynamic Lubrication Analysis

The proposed project is to extend the current research on lubrication analysis by means of a hybrid technique for global optimization. In a former NSC project (Distributed Genetic Algorithm for the Analysis of Thermohydrodynamic Lubrication, 2004/8-2005/7) a global optimization was successfully conducted in a computer cluster with standard Ethernet connection. Currently, a low latency cluster is being built to optimize the performance of the computing system (NSC project, 2007/8-2008/7). However, the computational demand is still large for a complex lubrication problem with multiple objectives and design variables. In a recent NSC project (The Metamodel for Numerical Fluid-Film Lubrication Analysis, 2006/8-2007/7) the author examines the usefulness of the metamodels in replace of the numerical models in the application of lubrication optimization. The results show that the metamodels can obtain a feasible solution almost instantly. Nevertheless, for an optimum design demands an accurate solution a time-consuming numerical calculation is usually required. In this project the author proposes an efficient global optimization technique, which uses a hybrid approach to accelerate the optimization process as well as to maintain the solution accuracy in a thermohydrodynamic lubrication analysis. The numerical calculation is to be conducted in parallel in a small computer cluster with low latency networking. The main issue is to setup an algorithm to replace the numerical models with a metamodel at some proper steps in the process of optimization. This will significantly reduce the overall execution time in solving a global search problem. Since a global optimization method is conducting search in many places simultaneously, the search can be accelerated further by using parallel computing. The global search methods to be investigated are genetic algorithms, DIRECT algorithm, and simulated annealing. The metamodels to be examined are artificial neural networks, Kriging method, and response surface method. The resulted hybrid scheme should be able to provide an optimum solution quickly without sacrificing the solution accuracy.

Project ID:PB9709-3567
External Project ID:NSC97-2221-E182-047-MY2