Sehr interessanter Artikel über den Beitrag von Machine Learning bei der Evolution der Aerodynamik des Porsche 919 Hybrid Evo.
This is when we decided to use some of the machine learning algorithms to help us first with the choice of the correct design parameters. Similar like in picture processing, one can use the database of available airfoils (in our case about ~1600 airfoils, m-selig.ae.illinois.edu) and train the machine learning models with these (so-called dimensionality reduction). After this analysis, it turns out, that something like an “airfoil DNA” can be extracted from all the data and that each of possible airfoils can be described only by a relatively small number of “airfoil genes” (only 5 to 10 depending on desired precision). Therefore, if an engineer is looking for an optimal airfoil, it is enough only just to vary these few “airfoil genes”.Once we could describe any possible airfoil with only just a few design parameters, the question was, how to find the best one. That is a typical optimization problem. Just to illustrate the complexity — if we describe each of the 2 airfoils with 5 design parameters, and if we also consider the relative position of the 2 wing segments and their size ratio as an optimization criterium, then together with the unknown flap angle in the opened position, we end up with altogether 15 design parameters. If we vary these design parameters only in e.g. 5 steps, calculate all the combinations and select the best one, we would need to analyze ~30 billion — each under multiple operating conditions. That is clearly not a feasible solution and still the resolution of only 5 steps per design parameter is extremely coarse.Therefore, also at this stage, optimization algorithms have to be used. Here again, we used the algorithms, that mimic the evolution process of the organisms in the nature and lead to the “fittest” one. Thanks to this, the optimum could be found only by analyzing hundreds to a few thousands of designs.The hundreds to thousands are already a relatively reasonable range, which can be simulated with the computational fluid dynamics (CFD) tools and with the use of high performance computing (HPC) resources available. Therefore, with the use of Machine learning for the airfoil description, Genetic algorithms for the optimization and computational fluid dynamic for the simulation on a HPC cluster, the optimal shape of the rear wing for the 919 Evo was found. Using conventional methods, billions of different shapes would have to be analyzed and even the best of these billions still would not reach the quality of the one, which was obtained by our implemented method.