Recently, many researchers have studied multi-fidelity meta-models to efficiently carry out design optimization using CAE. Single fidelity meta-model, built based on high
fidelity data only, usually requires ten times the number of design variables to reasonably approximate the exact model, whose computational burden can be heavy for the problems requiring long analysis time. However, a multi-fidelity meta-model is built based on the combination of low fidelity data and high fidelity data. In the multi-fidelity meta-model, high fidelity data are highly accurate but expensive to evaluate, while low fidelity data are cheap to evaluate but have a low accuracy compared to the high fidelity data. Multifidelity meta-model can efficiently approximate the exact model by using a small number of high fidelity data with a large number of low fidelity data. In this study, we studied the performance of a multi-fidelity meta-model with varying the number of samples and a variety of low fidelity models for eight mathematical problems. Furthermore we verified the efficiency of multi-fidelity meta-model from an engineering problem and a practical problem. From the results, increasing of the number of samples in our examples can improve accuracy of multi-fidelity meta-model and in variety of low fidelity models we can build up multi-fidelity meta-models as accurate as single fidelity meta-models. In an engineering problem we need high fidelity data, three times the number of design variables and low fidelity data, ten times the number of design variables as accurate as a single fidelity meta-model. In practical problem we build up multi-fidelity meta-models related to dust collection efficiency and pressure drop of cyclone. We can build up a multifidelity meta-model as accurate as a single fidelity meta-model with saving analysis time about 9.39% compared to a single fidelity meta-model in dust collection efficiency also in pressure drop we can construct a multi-fidelity meta-model as accurate as a single fidelity meta-model with saving analysis time about 28.98% compared to a single fidelity metamodel.