Journal of Computational Science and Technology Vol.4, No.1, 2010 Performance Map Construction for a Centrifugal Diffuser with Data Mining Techniques ∗ Koji SHIMOYAMA ∗∗ , Kazuyuki SUGIMURA ∗∗∗ , Shinkyu JEONG ∗∗ and Shigeru OBAYASHI ∗∗ ∗∗ Institute of Fluid Science, Tohoku University 2–1–1 Katahira, Aoba-ku, Sendai 980–8577, Japan E-mail: shimoyama@edge.ifs.tohoku.ac.jp ∗∗∗ Hitachi Plant Technologies, Ltd. 603 Kandatsu-machi, Tsuchiura, Ibaraki 300–0013, Japan Abstract Performance maps, which represent relations between performance and geometry pa- rameters, are essential for engineers to make a first decision on preliminary specifica- tion of a product to be designed. However, actual design often needs to consider var- ious performance and geometry parameters simultaneously. Therefore, the resulting performance maps must be constructed in a high-dimensional form. Based on these re- quirements, this paper proposes and demonstrates performance map construction with the aid of data mining techniques. Data mining can reveal characteristic patterns in high-dimensional data with performance and geometry parameters. Therefore, the data mining results make it easy to interpret complex features of performance vs. geome- try relations, and help engineers to discover new knowledge for engineering design through interpretation. The present demonstration of a centrifugal diffuser demon- strated that the data mining techniques are suitable and applicable to high-dimensional performance map construction, together with actual acquisition of new knowledge for diffuser design that was unknown from conventional quasi-one-dimensional nozzle theory. Key words : Data Mining, Computational Fluid Dynamics, Performance Map, Cen- trifugal Diffuser, Nozzle Theory 1. Introduction Recent progress in computer technology has brought about great developments in the field of simulation engineering. Simulation algorithms (formulation, modeling, schemes, etc.) are constantly improving in terms of efficiency, accuracy, and applicability. With such progress, numerical simulations can output various data, help engineers to discover new knowledge for actual engineering design, and promote further progress of simulation tech- nologies themselves. For knowledge acquisition, numerical simulations are expected to output as much data as possible. In actual situations, however, huge amounts of simulation data often lead to dif- ficulties at the data analysis stage. A major issue in data analysis is high-dimensionality. For example, computational fluid dynamics (CFD) simulations output the data of various physical properties (density, pressure, velocity, etc.) calculated at different spatial locations and differ- ent time iterations, and then evaluate various performances regarding fluid dynamics (output power, energy loss, noise level, etc.). Especially for fluid machinery design, CFD simula- tions are performed for different body geometries (shape, scale, topology, etc.). Thus, the output data from numerical simulations contain combination sets of various performance and geometry parameter values in a high-dimensional form. ∗ Received 18 June, 2009 (No. 09-0277) [DOI: 10.1299/jcst.4.36] Copyright c  2010 by JSME 36