Surrogate assisted EAs that use CFD analysis for fitness
determination have previously been used to design turbine
blades, finding interesting solutions with reduced computational
time [64]. Jin et al. [65] explored an ES with CFD
analysis to minimize the pressure loss of a turbine blade
while maintaining a certain outflow angle. The blade representation
used consisted of a series of B-spline control points.
The population was initialized with a given blade and two neural networks were used to approximate the pressure loss
and outflow angle, finding faster convergence than without the
surrogate models. Gräning et al. [6] used an ES with covariance
matrix adaptation to minimize the pressure loss of a
blade using 3-D CFD simulations. The ES was augmented
by a neural network surrogate model and used a preselection
resampling approach (where offspring are only generated from
individuals evaluated on the real fitness function), however,
significant improvement over a simple ES was not found.