Machine Learning for Aerodynamics - Deep Learning & Neural Networks applied to CFD simulations
HTML-код
- Опубликовано: 21 авг 2024
- For more information on adjoint shape optimization: • Aerodynamic Shape Opti...
In this video, we look at how machine learning / deep learning / neural networks can be applied to aerodynamic CFD simulations.
Neural Concept
We interviewed Pierre Baqué, CEO of Neural Concept, a Swiss startup developing & offering Deep Learning software. They have developed algorithms to connect 3D shape morphing, deep learning and aerodynamics.
senseFly
senseFly is a Swiss drone company that wanted to improve the flight time of their fixed-wing drones. senseFly, Neural Concept, EPFL (the technical University of Lausanne - École polytechnique fédérale de Lausanne) and AirShaper teamed up to apply Deep Learning to drone design to improve the aerodynamics, as improvements to the lift/drag ratio directly extend the range / increase flight time.
Deep Learning setup
The Neural Concept software can create & explore new 3D shapes to train its network, but it needs an aerodynamics component to give feedback on the lift/drag performance (and other aerodynamic parameters) of each design. For that, the Neural Concept software connected to the AirShaper cloud via an API interface.
Network Training
The training of the network was done in multiple phases with increasing accuracy. The initial warm-up of the network was done using older, in-house simulations from other projects. In the second phase, medium accuracy AirShaper simulations were applied. And in the final phase, high-accuracy AirShaper simulations were used for final tweaking of the network.
Output
Without any design input, the network came up with special drone shapes that partially matched what engineers had been applying for years in practice (anhedral/dihedral setup, ...). The lift/drag ratio was improved by more than 4%. Because the Reynolds number is quite different compared to large aircraft, so were the suggested design solutions.
AIRBUS
Neural Concept worked on the prediction of shock waves (transonic simulations). These results were presented at NEURIPS.
Future of Deep Learning for Aerodynamics
- Today
For industry specific, repetitive tasks, it pays off to train a network so that new designs can be analysed using the predictive model
- Short term
In the short term, machine learning can be used to make existing CFD codes faster and more accurate
- Long term
It's uncertain if it will ever work, but it might be possible to create generic Neural Networks that cover various industry segments, without needing to train the network.
----------------------------------------------------------------------------------------
Full Research Paper: www.airshaper....
-----------------------------------------------------------------------------------------------------------
The AirShaper videos cover the basics of aerodynamics (aerodynamic drag, drag & lift coefficients, boundary layer theory, flow separation, reynolds number...), simulation aspects (computational fluid dynamics, CFD meshing, ...) and aerodynamic testing (wind tunnel testing, flow visualization, ...).
We then use those basics to explain the aerodynamics of (race) cars (aerodynamic efficiency of electric vehicles, aerodynamic drag, downforce, aero maps, formula one aerodynamics, ...), drones and airplanes (propellers, airfoils, electric aviation, eVTOLS, ...), motorcycles (wind buffeting, motogp aerodynamics, ...) and more!
For more information, visit www.airshaper.com
That was a good interview. Waiting for more 👍
Thanks! For more info, you can download the research paper at airshaper.com/research
@@AirShaper I will thank you.
Great Info..next level of aerospace research.
Your videos should have million views . Great points on NN.
Thanks Daniyal!
The next step is using Quantum Processing to run the Simulations, then use AI to analyze the results to generate a new design and rinse-repeat, not to mention, I do believe Quantum processing is also Ideal for AI algorithms yes/no?
10/10
Thank you!
very interesting
Thanks!!
What constraints were set? Did it change static margins?
Dear Bill,
very few constraints were set and the main goal was the lift / drag ratio, as we were looking to increase the efficiency of the drone. In the process, it is likely that the static margin of the drone has changed a bit.
Can you please guide for CAD/CAE purpose
Hi Akhil, we don't provide guidance in terms of CAD/CAE - only CFD!
I am interested can you teach me how to connect with you.
can you present a demo with a software
Hi Veer,
do you mean a demo in which we show how the interface of the optimization software works?
You'll find more information on this website as well: neuralconcept.com/
Interested
If you wish to discuss a specific project, just contact us at info@airshaper.com!
Can you please guide me in CFD
We have a number of resources to get started:
airshaper.com/blog/best-aerodynamic-resources
The red on the nose makes me think the plane would benefit from a long beak like most birds have.
Indeed it would reduce the size of the red zone locally, but it would also increase pressure in other locations (as in the current design, the nose kind of "takes the hit" to reduce pressure elsewhere). I wouldn't be able to predict upfront if it would be better or not (and whether the beak of birds evolved to its current shape because of aerodynamics or something else). Also keep in mind that the Reynolds number could differ between these applications (meaning the flow regime might change and thus call for a different optimal shape). Thanks for the comment!
@@AirShaper Incredibly well thought out reply. I guess it would also change the Aerodynamics of turning and might make it less stable by affecting the airflow that reaches the ailerons. I was thinking of the Sears-Haack body, which I thought was better than using a secant ogive for optimal airflow. For simplicity, secant ogive is easier to model than Sears-Haack, but for less than the speed of sound, tangent ogive probably works just fine. i have always been obsessed with efficiency, but practicality has its merits too.
@@musikSkool Interesting to read that you're that deep into the topic! I guess some of your remarks also relate well to the uncertainty quantification that is being applied more & more to CFD, to take small (and perhaps large) variations in setup, geometry, ... into account to obtain a robust design that performs well over a large envelope!
With CFD being an "iterational" thing sometimes very far from reality, I would never rely on a machine learning. You just multiply errors put into program by people , instead of natural tests and experiments.
You could also apply machine learning to real wind tunnel data, but then it's very difficult to generate enough set points (you'd have to make hundreds of prototypes) and data (you would need a lot of pressure sensors). If you can first validate your CFD results versus wind tunnel data, you can then use simulations to train the network. But those indeed need to accurate enough for the problem you're looking at (which could mean you need to perform averaging, need a very fine grid, ...).
I will add that when using machine learning you don't always multiply the errors, but can average them out.
While training the networks you can have weights according to the accuracy of the CFD simulations, thus "trusting" only good samples, while using the "bad" ones just to interpolate in-between good samples.
The design phase can then be done with the ML algorithm and you can validate the last steps with high fidelity simulation/measurements.
Best method might be to calibrate your cfd by optimizing free parameters to match physical test data. Then train from the calibrated cfd