Note: I have since discovered I used the wrong viscosity for air on this experiment, so the results aren't valid for use on the earth. (Maybe Jupiter.)
A more intricate genetic algorithm that runs a 3D wind model and accounts for the cost of manufacture would almost certainly produce some more novel solutions that completely demolish our preconceived notions of what turbines should look like, just like the "jellyfish" cross-section. But it would also probably not effectively scale beyond the model size. Genetic algorithms almost always exploit any previously hidden little flaws you may have in your model simulator, such as the rounding error on floating point numbers at the limit of their precision, or even the way in which you represent airflows. They evolve to maximize performance in the modeling environment, so if your model is not perfect, the solution does not always translate to reality.
As such, you always have to test the evolved solutions in the real world before you can say they are better than the designed solution.
In fact I've thought about what the optimal way to do real world experiments is. Assuming you want to minimize them as much as possible, if making real prototypes is expensive.
I think you could use machine learning on the results of real world experiments to invent a model. It would predict how well each design would do. Then you could optimize the design to the machine learned model. Which of course won't be very accurate. But then it will give you a new test to do in the real world, which will give you data to make it even more accurate.
Additive volume-printer or subtractive CNC machinery would almost be required to make the physical test models.
Doesn't it seem to converge a little aggressively?
How do you feel about the potential bottlenecking pressures of a breeding pool of size... six?
The parameters of the genetic algorithm are so overwhelmingly ready put things permanently askew from random variations in early stages of the trial sequence that I'm not sure the sim really has a significant impact on the result at all. The author even seems to be aware of this, when he admits running it again does not generate similar results, and starting with a single piece of bias in the initial sample causes that result to dominate.
Genetic algorithms aren't something that can be thrown at a problem with zero attention to the parameters. Sorry, no algorithm is that magical.
Generally I think it's best to use a small population, or even just hillclimbing. But with the smaller population it should run and converge much, much faster. Then you can run it multiple times, and take the best result.
He talks about 20 minute tests, and then later uses a 12 hour test and gets a significantly different power number. I wonder how differently the evolution might proceed if he just had more power to run the full tests on each generation.