Aero engines are complex technical systems consisting of several coupled subsystems like fan, compressor, combustion chamber and turbines, each possessing a large number of design parameters to be determinated. This is usually based on numerical analysis where candidate designs are assessed by different models at some level of fidelity representing the engine subsystems. Routinely, a framework including e.g. evolutionary optimization strategies seeks for globally optimal designs. The downsides of using such an approach, however, are conflicts between the high number of design parameters and use of stochastic search methods on the one hand, as well as the high number of design evaluations required by evolutionary optimization and the costly design analysis causing improper evaluation time on the other hand. To overcome these issues, a common approach is to decouple optimization from analysis by using surrogates substituting original model-evaluations. These surrogate models like regression models are adapted to the original system input-output behaviour, trained by a relatively small number of samples drawn from the design space and analysed by the original model prior to the optimization. Evaluation of surrogate models is significantly faster than original model analysis, which in turn enables enough design assessments by evolutionary optimization strategies. However, this kind of approximation comes with drawbacks as well. To reach good approximation quality, the training of the surrogate models needs to be done with a sufficient number of costly training samples – especially in the case of a high-dimensional design space. Therefore, global sensitivity analysis may be used to cut down the design space to the most influential system parameters. An aspect, which has not been addressed in the literature, is the viability of designs caused by physical, numerical or other limitations, which are involved in the original analysis model. Thus, a trained surrogate model typically provides solutions for every design evaluation, even if it is not viable in the original design process. A subsequent optimization process, based on such surrogate models only, can be misled to optimized but inviable results. To address the problems mentioned above, this paper presents a combination of regression and classification models used as surrogate models for fast evaluating designs during evolutionary optimization, where the latter will estimate the probability of viability of a certain design. Two optimization criteria are defined regarding engine efficiency and weight, in particular specific fuel consumption SFC and mass m. For regression, various surrogate models are trained and tested to approximate both criteria on authentic sets of data sampled from an industrial aero engine design analysis tool. Then, different model-based sensitivity analysis methods are compared for reducing the dimensionality of the original high-dimensional design space. For classification, an analogous procedure is repeated to approximate the probability of a certain design being viable and reduce the dimensionality as well. Finally, both regression and classification models are integrated into an optimization framework to perform bi-criterion design optimization on a reduced design space while taking into account the viability of designs.