You are here

Surrogate Models and Mixtures of Experts in Aerodynamic Performance Prediction for Mission Analysis

TitleSurrogate Models and Mixtures of Experts in Aerodynamic Performance Prediction for Mission Analysis
Publication TypeJournal Articles
Year of Publication2015
AuthorsLiem, RP, Mader, CA, Martins, JRRA
JournalAerospace Science and Technology
Volume43
Pagination126-151
Abstract
Accurate aircraft fuel burn evaluation over a complete mission is computationally expensive, as it requires up to millions of aerodynamic performance evaluations. Thus, it is advantageous to use surrogate models as approximations of high-fidelity aerodynamic or aerostructural models. Conventional surrogate models, such as radial basis function and kriging, are insufficient to model these functions accurately, especially in the transonic regime. To address this issue, we explore several ways to improve the accuracy of surrogate models. First, we employ an adaptive sampling algorithm to complement a traditional space-filling algorithm. Second, we improve the kriging surrogate performance by including gradient information in the interpolation (a form of gradient-enhanced kriging—GEK), and by introducing a known trend in the global model component (kriging with a trend). Lastly, we propose a mixture of experts (ME) approach, which is derived based on the divide-and-conquer principle. We validate the developed surrogate models using aerodynamic data for conventional and unconventional aircraft configurations, and assess their performance in predicting the mission ranges by performing analyses on ten mission profiles. Our results show that the proposed ME approach is superior to the traditional models. Using a mixture of GEK models to approximate drag coefficients give us approximation errors of less than 5% with less than 150 samples, whereas the adaptive sampling fails to converge when training a global model. However, when we have a simple function profile, such as the lift and moment coefficients, using a conventional surrogate model is more efficient than an ME model, due to the added computational complexity in the latter. The range estimation errors associated with the ME models are all less than 2% for all the test mission profiles considered, whereas some traditional models yield errors as high as 20%-80%. We thus conclude that the ME technique is both necessary and sufficient to model the aerodynamic coefficients for surrogate-based mission analysis.
DOI10.1016/j.ast.2015.02.019
Citation KeyLiem2015