Configuring genetic algorithm to solve the inverse heat conduction problem

Abstract

Accurate design of heat treatment operations requires the knowledge of the Heat Transfer Coefficients (HTC), which quantity can be determined by performing Inverse Heat Transfer Calculations. The novel approaches for the estimation of HTC are based on heuristic optimisation methods, but the usage of these techniques raises several questions. In the case of genetic algorithms, there are not any rules of thumb for selecting the appropriate population size, mutation rate, stopping condition, and similar. The most efficient way to fine-tune these parameters is to run thousands of experimental tests and evaluate the results. However, in the case of inverse heat conduction, this has not been a viable option because of the high computational demand of fitness calculation which leads to a runtime of dozens of years. This paper presents a solution to this problem using a novel data-parallel direct heat conduction problem solver method implemented on multiple graphics accelerators. The ~100× speed-up achieved by this parallel algorithm made it possible to finish the necessary experimental tests in 15 weeks (instead of 29 years). Data gathered during these experiments are directly useful in practice. Based on these, it is possible to make recommendations for optimal genetic algorithm configuration parameters.