11 minutes ago, Bjoern said:

A bit of execution time penalty is totally worth not having to deal with the utter pile of BS that is anything C.

😁

10 hours ago, mSparks said:

cool toy 

https://paveldogreat.github.io/WebGL-Fluid-Simulation/

LOL that's sooo cool 😅

Ai-based CFD discussed somewhat recently on a YT channel who's got some publicly-acknowledged clout. Some of his interviewees are big names in the CFD world (Menter, Jameson). Overall summary: the industry is still skeptical. Especially aerospace. Jury is still way out on "foundational" CFD models which is what would be needed for a general-purpose consumer flight sim. Specialized models for specific geometry classes (e.g. 4door sedans) and boundary conditions still get a "maybe".

https://www.youtube.com/watch?v=aF140tpENN8

He uses the phrase "real-time CFD":

https://youtu.be/ayJKToUhp3Y?t=1390

 

CTO of Ansys (for now still the 800kg gorilla in the CFD world) discussing the complications:

https://youtu.be/9Ic5xgJt6BQ?t=2005

Certainly a foundational ai-CFD model would not be cheap to develop. Hundreds of thousands of high quality CFD simulations are estimated to be needed and that's an underestimate. Would someone like Ansys (who has the greatest chance of making something like this work) let that model out for free? Ansys products cost beaucoup $$. Real-time ai CFD coming soon to a consumer sim? Nope.

A random number generator is much cheaper and easier to develop than an AI CFD toolbox, while delivering the exact same results.

The videos are interesting, but they were focused on foundational models that can handle any geometry. The fact that foundation models are being research and worked on is amazing. Just a few years ago, many people thought it was impossible until the AI revolution. It's not a matter of if, but a matter of when for foundational models. However, the more interesting aspect as it relates to flight sim in the near term are specialized models that could predict the behavior of similarly shaped airframes. e.g. anything shaped like an airbus, or anything shape like a Cessna.

While the videos are focused on foundational models, I'm more interested in specialized models. Flight simulation does not require foundational models to take advantage of "AI-based CFD" and emerging technologies. We are not engineering the plane itself in flight sim. We are trying to replicate the flight characteristics of a known entity. Based on the surveys mentioned in the first video, even the experts believe that AI-Based CFD models are possible for specialized use cases, e.g. cars and planes. We've also seen other specialize cases earlier in this thread which I linked to.

Also as stated in the first video, a lot of the skeptics are now finally understanding what AI is actually capable of (not a fad). 

I can see a case where there is a model that is highly accurate at predicting the flight characteristics for certain types of airframes like airliners when trained on the appropriate data but might not be great for rotary wings if it wasn't trained on that dataset. In this case, a hybrid approach can be taken where developers could opt-in as is the case with MSFS current CFD implementation. In any case, AI-based CFD has the possibility the to be closer to reality than anything we could dream of today. After watching the videos, I'm even more convinced that AI-based CFD, particularly specialized models will be playing a major role going forward for rapid development, at least in the design iterations phases.

The only immediate limitation I see would be the data to train the models and the cost. There are already startups popping up though. AI-Based CFD coming soon to a consumer sim? Noone knows. It is impossible to say where the tech will be in 2 years or 4 years.

The industry does know that it's not happening any time soon for us plebs. As I alluded to months ago, the amount of CFD needed for even a specialized model is significant. And you still only ever have a statistical model at that. Carving up all those simulations to make a BET model would be the preferred path given the inherent precision of the method.