Investigation of Data Exploration Method for Thermal-Fluid Simulation Results Using Proper Orthogonal Decomposition

Abstract

A data exploration method based on POD was investigated to get clues to improve actual-product designs from a small set of rough-design candidates. The current method is based on the Oyama method which was used to acquire design knowledge of Pareto-optimal solutions obtained by evolutionary approach with large set of simulation models. The expanded method adopted (1) 2 steps POD analysis, and (2) voxel representation for different topology models. These approaches enabled to evaluate rough-design candidates which had non-parametric structures and those were difficult to handle in the original method. Especially, for the case of handling rough-design candidates, there was no explicit rule for model number ordering at the "first" POD analysis. We found that reordering the model numbers based on the first POD result was important, and rich information was obtained from the expansion coefficient vectors that made it possible to classify the feature regions and to detect feature regions with complicated flow structures at the "second" POD analysis. The effectiveness of the expanded method was verified in a cooling design problem for an optical disc drive. We found that the expanded POD-based data exploration method is a promising method for detecting key factors for actual product designs.