Sampling strategies for metamodel enrichment and automotive fan optimization

The strategy aimed at reducing the development time for a fan, which is oriented through the use of large meta-models which can be re-used and enriched along time. Efforts have been brought in the parameterization of the geometry and in the simulation process which provides pressure rise, torque and consequently efficiency. Numerical Designs of Experiments (DoE) are then conducted in an 11 factor problem to fill the space and to build a first kriging model. This latter assumes that the output is a gaussian field for which parameters are computed by maximum likelihood estimation based on the numerical runs. Uncertainties associated with any predicted values can then be assessed using the variance, and two small validation plans are used to measure statistically the errors. The variance given by the model is further used to map the areas in the domain which would need additional sampling. Then, two strategies are tested to select the most relevant sampling points, the first one being to reduce the range of the parameter variation, and the other one being to select them according to turbomachine design rules which would have dismiss some factor combinations. These two methods for sequential enrichment of the response are then compared and can even be combined with a trend given on the pressure rise. Answers from the kriging models are then assessed again in terms of statistical errors. These strategies will be described in the proposed paper through model comparison and optimization results.