Accurate aircraft fuel burn evaluation requires performing a detailed mission analysis covering the entire mission, from takeoff to landing. This process is computationally expensive, as it requires up to millions of aerodynamic performance evaluations, and thus it is advantageous to use surrogate models as approximations of the actual aerodynamic models. Training surrogate models is challenging due to the high nonlinearity of the aerodynamic performance functions in the transonic regime. Conventional surrogate models, such as radial basis function and kriging, are deemed insufficient to model these functions accurately. To address this issue, we explore several ways to improve the predictive performance of surrogate models. First, we employ an adaptive sampling algorithm in addition to the more traditional space-filling algorithm. Second, we improve the kriging performance by including gradient information in the interpolation (gradient-enhanced kriging), as well as by introducing a known trend in the global model component (kriging with a trend). Lastly, we propose a mixture of experts approach, which is derived based on the divide-and-conquer principle. In this last approach, we use multiple surrogate models as local experts to approximate different parts of the input space, using machine learning techniques to infer about the function profile to automatically partition the input space. These various surrogate models are tested using aerodynamic data for conventional and unconventional aircraft configurations. We then perform a surrogate-based mission analysis using the selected surrogate models. Our results show that the proposed mixture of experts approach can significantly improve the predictive performance when approximating the aerodynamic performance. For example, a mixture of five gradient-enhanced kriging models (with adaptive sampling) achieves 5% approximation error with around 100 samples, whereas the adaptive sampling fails to converge when training a global model. However, when we have a simple function profile, using a global model is more efficient than a mixture of experts, due to the added computational complexity in the latter.
展开▼
