Abstract
A two-step idea is presented for efficiently training a neural network (NN) surrogate of a parametric, high-fidelity, high-dimensional computational fluid dynamics model (HDM). It consists first in using a Gaussian process, an acquisition function, the HDM, and a scalar quantity obtained by post-processing the HDM to adaptively sample the parameter space. Next, a NN is trained for a quantity of interest using the HDM solutions computed at the sampled parameter points. This active learning approach is illustrated with numerical experiments for the prediction of the lift-over-drag ratio of a cambered NACA airfoil and the pressure coefficient of a flying wing aircraft mAEWing2 in large parameter spaces of flight conditions and shape design variables. The results demonstrate the superior efficiency and accuracy delivered by the proposed training over alternatives based on uniform and random samplings. The resulting surrogate models are suitable for time-critical applications such as design optimization and uncertainty quantification.