Has anyone seen a tool that can design custom RC airplanes from requirements?

[Realmdyz]

Someone introduced me to physicsx, but it wasn't suitable at all.

I am studying Artificial Intelligence.
I will discuss this with my aerospace friends
I hope we can build something
If you would like to help us too I would be happy
Thanks for your attention

Someone introduced physicsx to me, but it wasn't suitable at all.
do you see that before?

 

[telnar1236]

I remember an anecdote I heard back when I was in college - some big aerospace company or other had tried to get an AI to design an airplane by essentially giving it full control over the design and some target performance characteristics and invested a ton of effort in building the AI and training it and ran it on a supercomputer to give it all the power it could need. And then they set it running and it never came back with a good answer because there were too many variables and without some level of common sense/practical experience there's no good way to narrow things down to a workable option. Now this was pre-chat GPT and the technology is definitely better now than it was then, but I still suspect you'd run into the same issues

 

[telnar1236]

With not all that much guidance, I have gotten Gemini to give me workable dimensions, power systems, and construction methods for an RC plane, but it doesn't put it into CAD and I did have to convince it that all three gear struts should not go ahead of the CG to help the plane rotate, and I did need to know what questions to ask it in the first place - e.g. it would not have told me how much washout to include had I not asked it how much I should have
This is an example prompt that gave a fairly reasonable answer. I think the recommended CG is somewhat tail heavy but the rest seems at least plausible. The picture is very obviously wrong and has the EDF in front of the inlets

Prompt:
Could you please give me all critical dimensions required to build a 50mm EDF jet design of your choice? You may just list the airfoil types instead of needing to fully define them, but please give me dimensions like wingspan, length, wing sweep, position of the wings and tail, etc. Please also recommend a construction method and a set of commercially available parts I could use for electronics (receiver, battery, ESC, EDF unit, and servos). Finally, please generate a picture of the aircraft you have designed and described

Response:
Here is a comprehensive design concept, a dimension list, a construction guide, and a recommended electronics package for a 50mm Electric Ducted Fan (EDF) jet. I have designed this aircraft, the "Kestrel-50", specifically to be easy to build, forgiving for a first-time EDF pilot, and highly efficient.

I. Aircraft Concept: The Kestrel-50 Sport Jet​

The Kestrel-50 is a low-wing, single-engine sport jet. It is not a scale model of a real military aircraft, but rather a dedicated "park jet" designed for performance. Its defining features are a large, high-lift wing for slow-speed stability, a simple fuselage structure, and a highly efficient, direct ducting path for the EDF.

II. Critical Dimensions​

All dimensions are given in millimeters (mm) and are approximate guides for cutting your material.

General Dimensions​

Component	Dimension (mm)
Wingspan (Total)	700 mm
Fuselage Length (Total)	850 mm
Target All-Up Weight (AUW)	550g - 650g

Wing Dimensions​

• Airfoil: Clark Y (flat-bottom, high lift, modified slightly thinner for less drag).
• Root Chord: 180 mm
• Tip Chord: 110 mm
• Wing Sweep (Leading Edge): 20 degrees back.
• Wing Position: Mounted on the bottom of the fuselage. The Leading Edge (LE) of the root should be located 320 mm back from the nose tip.
• Dihedral: 2 degrees per wing panel (slight upward angle) for stability.
• Control Surfaces: Ailerons only (full span of the tapered section).

Tail Dimensions (Horizontal Stabilizer & Elevator)​

• Airfoil: Flat plate (6mm foam).
• Span: 280 mm
• Root Chord: 100 mm
• Position: Mounted on the top edge of the fuselage at the very rear.
• Control Surfaces: Single elevator (full span).

Tail Dimensions (Vertical Stabilizer & Rudder)​

• Airfoil: Flat plate (6mm foam).
• Height: 160 mm
• Root Chord: 120 mm
• Position: Mounted centrally on top of the rear fuselage/horizontal stabilizer.
• Control Surface: Rudder is optional for this weight class; elevons/elevator+aileron is sufficient. If included, make it the rear 30% of the surface.

Center of Gravity (CG)​

• CG Location: This is the most critical measurement. It should be 145 mm back from the root leading edge of the wing where it meets the fuselage. Always balance your model before flying.

III. Construction Method: "Foamboard-Skin" over "Profile-Spine"​

This hybrid method combines the simplicity of a profile plane with the sleek look and strength of a full-body model. It is perfect for a scratch-builder.

1. The Spine (Skeleton): Cut a vertical profile of the entire aircraft fuselage from one sheet of 6mm foamboard. This defines the length and height. Cut slots where the wing spar and horizontal tail will pass through.
2. Internal Formers: Cut circular or square formers from 6mm foam that define the cross-section of the fuselage (e.g., at the nose, cockpit, EDF intake, EDF exhaust). Glue these perpendicularly to the vertical spine.
3. The Wing: Build the wing from a 6mm foam base panel, scoring and folding a 3mm foam top skin over a simple spar (a 4mm carbon fiber tube is highly recommended). Use the flat bottom profile for easy building.
4. Ducting: Use smooth, flexible material like poster board, acetate sheet, or thin plastic rolled into a tube to create the intake and exhaust ducts. Do not use rough foam for internal ducting, as it creates too much drag.
5. Skinning: Once the electronics and ducting are installed on the spine, wrap the entire fuselage structure in 3mm foamboard or thin, heat-curled Depron skin.
6. Finishing: Tape all leading edges and control surface hinges with high-quality packing tape.

IV. Recommended Commercially Available Electronics​

Here is a balanced package of parts available from standard hobby retailers (like HobbyKing, Motion RC, or Amazon).

Component	Recommendation	Specifications
EDF Unit & Motor	Powerfun 50mm 11-Blade EDF (or FMS/E-flite equivalent)	Must be rated for 4S LiPo. Motor should be ~4300Kv to 5000Kv.
ESC (Speed Controller)	Hobbywing SkyWalker 40A / 50A Brushless ESC	A 40A ESC is sufficient, but 50A provides a safety margin. Must have a BEC (Battery Eliminator Circuit) to power the servos.
Battery (LiPo)	1300mAh - 1600mAh 4S (14.8V) 40C+ LiPo	4S is required for jet-like performance. Keep the capacity low to manage weight.
Servos	4x EMAX ES08A II (9g) Micro Analog Servos	These are plastic gear, very light, and reliable for small models.
Receiver (Rx)	Spektrum AR410 (DSMX) 4-Channel or manufacturer equivalent	Minimum 4 channels: Throttle, Aileron, Elevator, (Optional Rudder).

V. Design Picture​

Based on the specifications above, I have generated a visualization of the Kestrel-50 in flight, showing its final form, component layout, and specialized ducting design.

 

[telnar1236]

I also asked it to generate plans and what it came up with was surprisingly plausible for a profile jet if you ignore it having two horizontal stabilizers although definitely not the fancy fully 3 dimension plane shown

 

[Houndpup Rc]

I also asked it to generate plans and what it came up with was surprisingly plausible for a profile jet if you ignore it having two horizontal stabilizers although definitely not the fancy fully 3 dimension plane shown
View attachment 258652

Also, it look like the pieces are a touch out of proportion compared t other pieces..But better than I expected!

 

[telnar1236]

Also, it look like the pieces are a touch out of proportion compared t other pieces..But better than I expected!

Yeah, not usable, but not that far from it

 

[Houndpup Rc]

Yeah, not usable, but not that far from it

Yep! Surprising close compared to what I have gotten in the past with ChatGPT!

 

[Realmdyz]

I also asked it to generate plans and what it came up with was surprisingly plausible for a profile jet if you ignore it having two horizontal stabilizers although definitely not the fancy fully 3 dimension plane shown
View attachment 258652

The designs you got are good but still have some issues. I think if a professional aerospace engineer (according to what I talked to my friend) can supervise these, the output we have will be almost ready to fly, plus the aerospace engineer can also produce the CAD file for it. The difference between the example you gave and what is currently the case is that currently GPT and things like this are halfway there and we just need to complete this path. I will be back soon with the final product.
Thanks for your comments, it will definitely be useful.

 

[telnar1236]

The designs you got are good but still have some issues. I think if a professional aerospace engineer (according to what I talked to my friend) can supervise these, the output we have will be almost ready to fly, plus the aerospace engineer can also produce the CAD file for it. The difference between the example you gave and what is currently the case is that currently GPT and things like this are halfway there and we just need to complete this path. I will be back soon with the final product.
Thanks for your comments, it will definitely be useful.

Don't think you even need an aerospace engineer - most RC plane design is common sense and you get people from all walks of life on this forum. Learning CAD takes time, but anyone can do it so if you want to get into designing planes, don't let CAD hold you back. However, I think supervising AI will generally take longer than just designing from first principles with where the technology is at.
Supervising Gemini, and translating the output into CAD myself, I could get this - a totally flyable design for 3D printing.

But not using AI and instead taking the approach of designing based on research into what had already been built and first principles, I was able to create each of these three designs in less time and with less effort. These all use the same power system and are more or less the same size, but the AI design would top out around 75 mph (projected) vs. 133 mph, 104 mph, and 82 mph (projected) respectively.

 

[L Edge]

The designs you got are good but still have some issues. I think if a professional aerospace engineer (according to what I talked to my friend) can supervise these, the output we have will be almost ready to fly, plus the aerospace engineer can also produce the CAD file for it. The difference between the example you gave and what is currently the case is that currently GPT and things like this are halfway there and we just need to complete this path. I will be back soon with the final product.
Thanks for your comments, it will definitely be useful.

Is there a tool to design everything?, the answer is no.

AI is in it's infancy and a lot of info it searches for comes from group's like FT, RCGroups etc., as well as the real design world of aviation. So if you do not provide enough info in the input, the search output will "fly, to it will crash" when built. Just like the design above, it will need to be "tweaked" to fly. Especially EDF's.

AI can help you lay the foundation of your RC plane, but it may end up given you a tall building rather than your castle.

Only when I can cross check AI and some other source(white paper, thesis) do I use it in designing a plane.

 

[Realmdyz]

Don't think you even need an aerospace engineer - most RC plane design is common sense and you get people from all walks of life on this forum. Learning CAD takes time, but anyone can do it so if you want to get into designing planes, don't let CAD hold you back. However, I think supervising AI will generally take longer than just designing from first principles with where the technology is at.
Supervising Gemini, and translating the output into CAD myself, I could get this - a totally flyable design for 3D printing.
View attachment 258670
But not using AI and instead taking the approach of designing based on research into what had already been built and first principles, I was able to create each of these three designs in less time and with less effort. These all use the same power system and are more or less the same size, but the AI design would top out around 75 mph (projected) vs. 133 mph, 104 mph, and 82 mph (projected) respectively.
View attachment 258671
View attachment 258672
View attachment 258673

Thanks for your advice and comments, I have a question!
What problems do you feel AI tools currently have for designing an RC plane?!
At first I felt the problem was that they don't provide CAD files, but you say that this is not the problem
I would appreciate it if you could tell me what their main problems are (the more complete you tell me, the happier I will be and the stronger we will be in designing our tools)

This is my question to the entire community and I would be grateful to anyone who comments, especially the other people who have helped us so far.

 

[Realmdyz]

Is there a tool to design everything?, the answer is no.

AI is in it's infancy and a lot of info it searches for comes from group's like FT, RCGroups etc., as well as the real design world of aviation. So if you do not provide enough info in the input, the search output will "fly, to it will crash" when built. Just like the design above, it will need to be "tweaked" to fly. Especially EDF's.

AI can help you lay the foundation of your RC plane, but it may end up given you a tall building rather than your castle.

Only when I can cross check AI and some other source(white paper, thesis) do I use it in designing a plane.

Thanks for your comment.
The question that comes to my mind is what do you see as the problems that you see in AI right now that if solved we would have a perfect RC design?

 

[Realmdyz]

Yep! Surprising close compared to what I have gotten in the past with ChatGPT!

I would be grateful if you could answer.
The question that comes to my mind is, what do you think are the current problems in AI that, if solved, would lead to a perfect RC design?

 

[telnar1236]

I think you can categorize issues with AI into 3 main areas, hallucinations, lack of physical intuition, and consistency.

Hallucinations are a broad issue with AI models and are especially prevalent in areas where the subject is fairly niche which is true of building RC planes. As an example, I asked Gemini to design a speed plane with a top speed around 130 mph (I've been working on this recently so have a fairly good sense of exactly what it should be outputting). It spent the next few responses trying to argue that I need a 45-degree swept wing to delay compressibility issues which don't typically appear till upwards of 400-500 mph and were completely irrelevant as design considerations. Honestly, this strikes me as the biggest issue since if you can't trust the outputs, then no one will use the tool.

A lack of physical understanding/spatial awareness means that AI is also unable to catch any errors. Any person building a plane will make the occasional mistake, but it's pretty obvious to a person if the propeller is on backwards, or the motor is spinning the wrong way, or if the CG is way off and the plane tips over, but this is not obvious to an AI model. My favorite error I've gotten asking AI to design a plane remains it telling me to put all three gear struts in front of the CG so the plane rotates easier for takeoff. Basically, it makes mistakes that a person with our intuitive understanding of the real world would never make and has no way of knowing that those are in fact mistakes.

And finally, consistency in what it outputs remains an issue. Because AI lacks a mental model of what it's physically trying to build it has a tendency to output a bunch of design elements that would all be fine on their own, but which don't work well together. If you ask it to draw a picture of what it designs, it also doesn't reflect the actual dimensions it gives. For example, the pretty sport jet shown in the pictures for the test prompt I showed earlier in the thread are quite different from what the dimensions listed in that response give you. This is what those dimensions would result in:

vs. the pictures:

Either could be developed into something flyable, but they're quite different and therefore you couldn't base parts of your design for one on the other. Since these came from the same prompt asking for a single design, that's a problem

 

[Houndpup Rc]

I think you can categorize issues with AI into 3 main areas, hallucinations, lack of physical intuition, and consistency.

Hallucinations are a broad issue with AI models and are especially prevalent in areas where the subject is fairly niche which is true of building RC planes. As an example, I asked Gemini to design a speed plane with a top speed around 130 mph (I've been working on this recently so have a fairly good sense of exactly what it should be outputting). It spent the next few responses trying to argue that I need a 45-degree swept wing to delay compressibility issues which don't typically appear till upwards of 400-500 mph and were completely irrelevant as design considerations. Honestly, this strikes me as the biggest issue since if you can't trust the outputs, then no one will use the tool.

A lack of physical understanding/spatial awareness means that AI is also unable to catch any errors. Any person building a plane will make the occasional mistake, but it's pretty obvious to a person if the propeller is on backwards, or the motor is spinning the wrong way, or if the CG is way off and the plane tips over, but this is not obvious to an AI model. My favorite error I've gotten asking AI to design a plane remains it telling me to put all three gear struts in front of the CG so the plane rotates easier for takeoff. Basically, it makes mistakes that a person with our intuitive understanding of the real world would never make and has no way of knowing that those are in fact mistakes.

And finally, consistency in what it outputs remains an issue. Because AI lacks a mental model of what it's physically trying to build it has a tendency to output a bunch of design elements that would all be fine on their own, but which don't work well together. If you ask it to draw a picture of what it designs, it also doesn't reflect the actual dimensions it gives. For example, the pretty sport jet shown in the pictures for the test prompt I showed earlier in the thread are quite different from what the dimensions listed in that response give you. This is what those dimensions would result in:
View attachment 258690
View attachment 258691
vs. the pictures:
View attachment 258692
Either could be developed into something flyable, but they're quite different and therefore you couldn't base parts of your design for one on the other. Since these came from the same prompt asking for a single design, that's a problem

Agreed! I did some research last night and what might be a viable system is to train your own AI system and you would manually design a bunch of plane and program the specs and 3D models into it..
After say 15 different types you could ask for a model and it would make it similar to existing models but with the shape you want ect, then you could take a FC, put it in that model that it made fly it then load the flight data back into the AI system so it can "learn" form its designs. As far as the 3D modeling part i think Freecad might work, because you can model stuff in that with just python script.

 

[telnar1236]

Agreed! I did some research last night and what might be a viable system is to train your own AI system and you would manually design a bunch of plane and program the specs and 3D models into it..
After say 15 different types you could ask for a model and it would make it similar to existing models but with the shape you want ect, then you could take a FC, put it in that model that it made fly it then load the flight data back into the AI system so it can "learn" form its designs. As far as the 3D modeling part i think Freecad might work, because you can model stuff in that with just python script.

I think part of the issue is also that we use AI as a blanket term for multiple different types of algorithm. We use it to describe both LLMs which are not great at engineering and other types of machine learning which are very useful as engineering tools but aren't particularly user friendly and tend to need to be coded specifically for the job you want them to do. Tools like what Houndpup Rc is describing are great for optimization of a general design layout, but you need to operating firmly within the design space set out by your training data and with specific numbers as targets for what you want your plane to hit. Here's a good video where someone talks about using this kind of machine learning tool to optimize a flying wing:

It also reveals the issues with AI optimization - just looking at the design, I can point to some issues with the shape of the trailing edge where it joins the fuselage that could be fixed to notably reduce drag, but because the AI model was not given the ability to optimize that, it didn't have the understanding to even try to look at it

 

[L Edge]

Pose question to AI:
Edge540 with 55cc motor ,rcplane, lost left main wheel in flight, explain in detail, 2 different ways to land the plane undamaged, understand yaw,pitch,roll.
Summary::
First way, screwed up not using Rudder at all. 2nd way, land it belly up!!!!! in grass . Design by AI, don't think so.

It is like the into of microcomputers, I had 1 K of memory to write Basic. Wait for a number of years.

 

[telnar1236]

Pose question to AI:
Edge540 with 55cc motor ,rcplane, lost left main wheel in flight, explain in detail, 2 different ways to land the plane undamaged, understand yaw,pitch,roll.
Summary::
First way, screwed up not using Rudder at all. 2nd way, land it belly up!!!!! in grass . Design by AI, don't think so.

It is like the into of microcomputers, I had 1 K of memory to write Basic. Wait for a number of years.

Which AI model are you using? I've generally had reasonably good results with Gemini so long as you're not expecting a fully fledged ready to go design. In this case, using the same prompt, I think it gave a broadly reasonable answer. I'd probably try to touch the tail wheel before the left strut but that doesn't necessarily make the answer bad or wrong

I like the analogy to an old computer though - definitely at similar stages in the development of the technology. Remains to be seen if generative AI is a PC or a Palm Pilot though

 

[L Edge]

Which AI model are you using? I've generally had reasonably good results with Gemini so long as you're not expecting a fully fledged ready to go design. In this case, using the same prompt, I think it gave a broadly reasonable answer. I'd probably try to touch the tail wheel before the left strut but that doesn't necessarily make the answer bad or wrong
View attachment 258702
I like the analogy to an old computer though - definitely at similar stages in the development of the technology. Remains to be seen if generative AI is a PC or a Palm Pilot though

As I explored AI, I found that most of the data comes from Rc Groups, FT, etc., which limits possibilities. In fact your name, telnar123's data and statement shows up if you asked it the right way.
If you remember, we discussed end cones over motors, vane straighteners and my adding tape to reduce the gap between blade and wall. AI found a MIT white paper tested a VTOL plane(4 passenger) EDF blade gap and came up with a formula to maximize Static thrust. Point made was they still need to test for the max dynamic thrust in flight (reducing exit) to see what happens with the inlet escaping EDF gases mixes with the inlet flow. Now we have an additional issue that needs to be looked at.
Now you can see where AI can't really design a RC plane at this stage of the game.

 

[Mr Man]

I really don’t see AI being able to design a plane anytime soon. Like @telnar1236 and @Houndpup Rc have said; there are just way too many variables, and AI really isn’t that good at designing stuff