An engineer wants to improve the design of an existing fixed-wing UAV by increasing lift by 1.5 times in order to let it carry heavier loads. He wants to leverage on drag measurements for 10 previous prototypes which were tested in a wind tunnel, on a past simulation results database produced by its department over the years with 500 simulated shapes. He can also use new simulations, but each call to a simulation costs 100$ and takes one day on a cluster of machines.
Step 1 – Preliminary design
The engineer needs to show a prototype to its management team and quickly produce a good design which has exactly the right behaviour. He feeds its historical data into our Neural Network in the form of 3D surface meshes (standard “.stl” or “.obj” files) and lift values. Our network will be able to quickly create a approximate model for the lift of an arbitrary new design. Our network will be able to quickly create an approximate model for the lift of an arbitrary new design. This lets our neural network explore very quickly all possible designs and provide a first drone shape which approximately meets the target criteria.
Contrarily to existing “surrogate methods”, which also create an approximate model based on ML to quickly estimate drag, our software can leverage past data from many different sources, without them having an underlying shared parameterization. This is critical in the case of a company which has a large legacy of simulation results.
Step 2 – Manual search for a good design
The engineer now attempts to manually modify this initial design to improve it. The traditional option would be to iteratively try a few designs and test them in simulations. However, once again each of these iterations takes a full day of computation, which slows down work drastically. Thanks to our solution, he can get an estimate of the lift of each new design, interactively on its CAD tool. Since our algorithm also provides a confidence score on the lift prediction — i.e. how close it is expected to be compared to what the simulator would give- he is wisely advised to call back the simulator when necessary and this new data point will be used to refine the estimates.
Existing methods based on simplified flow equation can also provide an approximate simulation relatively quickly. However, it necessitates to parametrise an approximate simulator, which will in most cases return results which are completely different, even for the known cases, from those of the actual simulator. On the opposite, our network will always provide results which are close to the ones of the simulator.
As opposed to parametric surrogate methods — e.g. Kriging –, our approach leaves the user completely free in the design. He can use his own CAD software, import designs from outside.
Step 3 – Numerical shape optimization
The work that was done manually by the engineer in the above case, can be automatised and accelerated. Hence, in order to maximise the performance of its design, the engineer now uses our automatic shape optimization loop. As in step-1, the engineer provides a set of design goals. He chooses to interface our program to the simulation software he was using — we are already compatible with several popular ones —, which will be called back only when it is necessary to improve the accuracy of the prediction. This leads to a design a human could not have thought of, bringing performance gains and time savings. With the use case showed here, we managed to achieve a design with 20% better performances. We are still collaborating with the client with further objectives to fulfil.
Previous methods for shape optimisation based on response surfaces and implemented on commercial softwares, are limited by the fact that they need a low dimensional parametric representation of the shape. With our approach, the optimisation algorithm can create arbitrarily complex shapes, and the user can easily manually interact with the optimisation algorithm. Therefore, numerical optimisation becomes much more smoothly integrated in the whole engineering process.