Understanding everything in this video is NOT necessary to understand how to read EM diagrams. This is just a documentation of the processes I used to generate the graphs and for those interested in trying it themselves (I do NOT recommend) Audio and presentation quality was also a bit rushed so i hope you enjoy random ASMR with your whiteboard thunder
I stand corrected - THIS is why I dropped out of working out my own EM diagrams. Maths. Your dedication is incredibly commendable - with your maths, maybe I'll finally start making some graphs... 😅
"Run while you can", "understandable that this isnt the most interesting thing" - hell naw thats why I'm here, seeing maths, physics and data science applied to some video game like war thunder is awesome! Great video and explanations.
This certainly isn't the average war thunder player's idea of an interesting video, but I'm glad that you find this stuff interesting like i do! I'm working on another technical video so look forward to that
well i dont think there are other people doing things exactly the way im doing them. Most other people are working on this analytically and while they've made progress, it's not really presentation worthy yet since its not very accurate yet. I'm not completely done with this yet either, consider this more a work in progress update for yall lol
After watching both parts twice, I have determined that my math comprehension is too low to fully understand this. The good thing is that your explanations are very concise, so it won't take long for that.
I think part of the reason its a bit hard to grasp is that I haven't shown any application yet. Im working on a part 3 video where I go into comparing these graphs so that might help ground all the math mumbo jumbo
WTRTI made the SEP calculation better as of v2.3.3, although I know you probably already know about (or maybe had a hand in) it I loved the video! thanks for putting in effort into a project like this
yeah, it seems he followed through with the suggestions i made a few months ago. I haven't tested them to confirm that they work though, and since i already have a homebrew program to calculate the stuff, I think I'll stick to it for the sake of consistency going forwards
glad you found it interesting! tbh I don't think most war thunder players are that interested in this kind of stuff, its always really validating to see that yall like watching whiteboard and math thunder lol. One thing I think about is how much personality/haha funni i should put in a video for those with shorter attention spans but for technical videos like this i think its probably better I leave that out since they're not likely to watch it anyways lmao
Really in depth vid, obviously it took a lot of effort. I liked that you showed where you are on the EM diagram in maneuvers via that blue trace around 5:30, makes reding EN diag. more intuitive. Next time comparing 2 fighters with such diagrams would be nice. For example by subtracting one diagram from another to obtain a EM diagram of differecnes between 2 planes. So you calculted TAS based on IAS by using air density only to obtain more precise TAS than in localhost or WTRTI? wow In 6th order polynomial, if you hadn't removed 1st order of y terms and odd order terms, wouldn't the fitting process just put very low 'a' coefficients to these terms? Then the surface would end up quite similar, but less precise, right? Did you use least squares regression to fit the model to data? About flight tests - you could use simplified sim controls and turn with mouse joystick. You could remove rudder input in mouse joystick settings, and make aileron input weak so that mouse joystick only uses elevator and a bit of alieron.That way it's easier to make precise constant alt turns that way, but it's still tough.
1) Thanks! I felt like I could've made it more in depth but for most people I think this was a good enough tour of what methods I used to get the plots. And for people much more in the weeds (like you), they could probably already follow along. 2) Thanks! I was wondering how I would demonstrate the testing regime. Might've been better if I had animated the lines to move with the plane but within my whiteboard skillset I think it was a decent enough compromise 3) Yep, that's the next step. I still need to write up a 2d comparison plotter, which is gonna be annoying lol. I genuinely spent more time writing up the plotters than doing the actual math and fitting cuz idk how to use matplotlib and plotly very effectively. 4) Yeah I went precise nosevelocity --AoA,AoS--> precise IAS --densityratio--> precise TAS. It was 6 months ago when I was takling to AVB on his discord that i realized WTRTI has too many problems and subjective choices made in terms of how its calculations are done that I decided to just stick with the localhost numbers and work from there myself. Very useful logging tool but lags the crap out of my computer and I can't trust the calculated numbers it spits out. 5) It might, but it also might not. In the previous two videos you can see that the old EM Diagram had contours that weren't perpendicular to the x axis. Obviously this is moreso a problem with the lack of accurate and precise data in the midsections but I also don't want to spend hours filling in the empty space in the middle. For the purposes of Air RB, the most important sections are the not-turning and fully-turning anyways, the middle is just sorta there as filler. Of course it would be nice if it were 100% accurate but because the method is inherently messy, I was always just kinda looking for a "close enough". For a more accurate picture, you'd probably need to just start from the other end and solve flight models analytically. I used scipy.optimize.curve_fit, and I'm not really that familiar with it, but when eyeballing the 3D surface plot with the scatter data it looked like a sufficient enough fit. 6) Probably... but I'm also not really that comfortable with sim controls, and I wanted the data to be well-fit for air RB. You can see the scatter data is mostly clustered around the not-turning and fully-turning lines so the fit works best there, which is what I want when comparing air rb performance. If I were using sim joystick controls, I would have to consider things like making sure not to stall, and I didn't want to deal with that. Also I think the curve fit surface came out pretty nicely considering everything so I'm not too fussed about a bit of rolling and yawing. keyboard elevator seemed to work pretty okay. If anyone were to check out the code then it'd be you lol, its uploaded in the gdrive that I've shared in the description so you could check it out if you wanna
I can see why it's powerful, but i am a plotly noob. maybe if i watched some dedicated tutorials on it but im only using it cuz 3d matplotlib sucks booty
Yea, the vid is good! It might be too advanced if one won't search things up and rewatch, but that's not always a bad thing. If you'll need help with plotly I'll gladly help, tho chatgpt is likely better.
yeah chatgpt is usually okay at pointing me in the right direction tho it screws stuff up too often to be used directly but it tells me what libraries are worth looking into the documentation for
Great work! I don’t have any real complaints today really good job. I definitely wouldn’t have expected that parametric fit to work so well, but honestly it looks pretty great! You could always look at trying some simple interpolation or knn algorithms. A final thought about overcoming some of these limitations would be to try and build a model using test data in order to implement the analytical method. Its likely gaijin used reasonably simple models in thier drag calculations so that you can probably get a reasonable drag Polar without too much hassle by varying aoa at various speeds and comparing to simple models. For what it’s worth if you can get a good model of drag, the engine data is pretty accessible in the dataleaks, a couple years ago i put together a tool for parsing it. Id also be happy to pass along some info on drag modeling from my aircraft design courses
Thanks! I was considering using a hybrid testing/analytical method. I'm fairly confident I know how drag is calculated atleast for singular parts of the aircraft in gaijin's war thunder. I might pursue it in the future but for now I think I'll stick with this until it becomes too unweildy/messy
@@CatWerfer as long as you know the functional form you shouldn’t need to worry about any part buildup, you can just approximate the whole-aircraft drag Polar and use it to recreate all of the different EM charts, thats how they originally did it aswell was just taking drag polars from the engineering groups and validating with flight tests. I think your approach would be very similar to what you’ve just done. This is perfectly fine for props itll only matter with supersonics due to the limitations you already laid out. Either way excited to see your progress!
4th order didn't work as well, was a bit too smooth and there were regions of visible over and undershoot. 6th is much better and there are no visibly eyeball-able deviations from the scatter data. Yeah I'll try to think about how I'd approach a hybrid experimental/analytical method for the future but for now I think I'll wrap up this series with this. I was planning on it only being a 2 parter but this has gotten much longer than i thought it would be
For more accurate testing without instructor spikes i would recommend setting up mouse joystick with autotrim, set joystick to square change from simple to standard to allow full motion and remove deadzone and make it not control rudder. Instructor kills turning on some planes and will make results look bad when flight model is different.
Someone else suggested this too but for what it's worth it don't think I care enough to fill out the middle accurately to do sim controls cuz at that point I have to add in a whole other set of restrictions like not going over stall aoa and I'm only really interested in edge behavior for air rb. That probably makes it so these graphs apply less in sim but if the performance is gimped in air rb then I want that to be reflected
Regarding the testing methods: it might be better to use full-real controls/mouse-as-joystick as you won’t have to fight the virtual instructor when you want to find data with only one rotational axis of input. Even with keyboard-only inputs, the virtual instructor still sometimes limits the maximum deflection of control surfaces depending on the air speed and AOA (One example being recovering from a stall with only keyboard inputs)
@lukeskywalket2894 I'm an air rb players so the instructor limits are important to me. Just because the 109 can pull better in sim doesnt mean than it does in rb, and i care mich more about rb performance than for sun. Using sim controls would also mean I'd have to worry about not stalling in flight and I'm also not used to the attached camera view that sim controls puts you in. If you want to test for sim you can feel free to use my code which is in the shared drive
@@CatWerferah, that makes sense. I’d still recommend using one of the simplified mouse-as-joystick control modes though, since there are ways to get more fine-tuned inputs while still having the limitations of the virtual instructor
This method of calculating SEP should be the default in WRTI I suggest you try the F8F-1 next for comparison, there is a lot of talk about its flightmodel recently with the drop to 4.7 It bleeds lots of energy with the slightest aoa Next step is to use these to verify if the flight models are accurate, though gaijin will drag their feet no matter what
The reason they filter and smooth data for wtrti is because its primarily meant to be an addiitoonal hud element and it would be annoying to see a number jittering all over the place. Most people arent gonna care about infinite precision for sep and those who are can prob calculate it themselves, though it would be much nicer if they just made an unfiltered option available. I find the precsision cutoff to be more annoying tbh. Tbh I don't care too much about realistic flight models too much as long as they don't feel too horrendous to fly. I'm also not really planning on using this to crosscheck flight models since this only reflects air rb performance anyways
@@CatWerfer It's already very jittery and hard to read, especially during climbs unless you decrease the polling rate to 1000ms. I suspect this is due to the low precision of using TAS. I suspect many meta-defining and/or uncompetitive FMs are the way they are because they are badly modeled (think P-39, 152H, A6M5, or F8F, F6F, 190A5) Gaijin has had plenty of turnover in its FM team in the past decade, FM quality standards are likley all over the place. Such a critical aspect of the game should be taken more seriously
@KekusMagnus i agree, I've suggested changes to the developer again but he seems like he's got more important things om his plate right now. Yeah flight models are off and it kinda sucks that gaijin is so horrible at balancing but if gaijin dismisses actuakly valid evidence I don't think this will change much until we can get more complete and accurate data
yeah it is possible since you can log pitch and roll angle of the aircraft as well, it would probably give a more accurate turnrate. However for the purposes of comparing aircraft i think what i have here is probably alright
Do you know the TH-cam channel "Gregs airplanes and automobiles" ? Its awesome, its like a university class about aviation, focusing in WW2. It studies in great detail the performance of WW2 aircraft including EEM charts, engine technology performance evolution, human factors in cockpit design, etc. Greg has like 8h of content just about the p47 alone. His content leaves me with my jaw dropped. If you made this video you will like his channel for sure!
Yeah I saw some of their stuff, really high quality! I might try something like that if I have the mental fortitude to spend 8hours talkinga bout one plane lmao
Understanding everything in this video is NOT necessary to understand how to read EM diagrams.
This is just a documentation of the processes I used to generate the graphs and for those interested in trying it themselves (I do NOT recommend)
Audio and presentation quality was also a bit rushed so i hope you enjoy random ASMR with your whiteboard thunder
I stand corrected - THIS is why I dropped out of working out my own EM diagrams. Maths.
Your dedication is incredibly commendable - with your maths, maybe I'll finally start making some graphs... 😅
@@Adrella_Highwing code is in the drive folder if u wanna take a look
"Run while you can", "understandable that this isnt the most interesting thing" - hell naw thats why I'm here, seeing maths, physics and data science applied to some video game like war thunder is awesome!
Great video and explanations.
This certainly isn't the average war thunder player's idea of an interesting video, but I'm glad that you find this stuff interesting like i do!
I'm working on another technical video so look forward to that
Babe wake up sweatlord tutorial version 2 just dropped
I'm sliding around on my belly like a penguin with all this sweat, you cant catch me
Cooked hard on this one.
i think i burnt the kitchen down
@@CatWerfer only the laptop
lol i had to shut down my computer for 4 hours to cool it off while i was testing once
@@CatWerfer i have been there man. Much respect
just wow, thanks for sharing the whole thing, there are many who were on this but they didn't bother to share, thank you very much!
well i dont think there are other people doing things exactly the way im doing them. Most other people are working on this analytically and while they've made progress, it's not really presentation worthy yet since its not very accurate yet. I'm not completely done with this yet either, consider this more a work in progress update for yall lol
After watching both parts twice, I have determined that my math comprehension is too low to fully understand this. The good thing is that your explanations are very concise, so it won't take long for that.
I think part of the reason its a bit hard to grasp is that I haven't shown any application yet. Im working on a part 3 video where I go into comparing these graphs so that might help ground all the math mumbo jumbo
Someone said i had the voice of a 14 year old and with the voicecracking in this video i am not beating those allegations 😭
Add some comfy long socks and it would be just in line with a typical warthunder enthusiast 😂
@@darthcalanil5333💀
you really went hard on this one. hope you get an ultra mega gaming pc to run this shit on. awesome work my friend 👍🤝
I don't need a gaming pc i just need WTRTI to fix its memory leak issues 😭
Man I love data processing (till I can leave)
My god, the headache is back but this is so fascinating!
WTRTI made the SEP calculation better as of v2.3.3, although I know you probably already know about (or maybe had a hand in) it
I loved the video! thanks for putting in effort into a project like this
yeah, it seems he followed through with the suggestions i made a few months ago.
I haven't tested them to confirm that they work though, and since i already have a homebrew program to calculate the stuff, I think I'll stick to it for the sake of consistency going forwards
Great work, I watched the whole thing even though I've quit war thunder years ago.
glad you found it interesting! tbh I don't think most war thunder players are that interested in this kind of stuff, its always really validating to see that yall like watching whiteboard and math thunder lol.
One thing I think about is how much personality/haha funni i should put in a video for those with shorter attention spans but for technical videos like this i think its probably better I leave that out since they're not likely to watch it anyways lmao
Really in depth vid, obviously it took a lot of effort.
I liked that you showed where you are on the EM diagram in maneuvers via that blue trace around 5:30, makes reding EN diag. more intuitive.
Next time comparing 2 fighters with such diagrams would be nice. For example by subtracting one diagram from another to obtain a EM diagram of differecnes between 2 planes.
So you calculted TAS based on IAS by using air density only to obtain more precise TAS than in localhost or WTRTI? wow
In 6th order polynomial, if you hadn't removed 1st order of y terms and odd order terms, wouldn't the fitting process just put very low 'a' coefficients to these terms? Then the surface would end up quite similar, but less precise, right? Did you use least squares regression to fit the model to data?
About flight tests - you could use simplified sim controls and turn with mouse joystick. You could remove rudder input in mouse joystick settings, and make aileron input weak so that mouse joystick only uses elevator and a bit of alieron.That way it's easier to make precise constant alt turns that way, but it's still tough.
Also, Plotly FTW!
1) Thanks! I felt like I could've made it more in depth but for most people I think this was a good enough tour of what methods I used to get the plots. And for people much more in the weeds (like you), they could probably already follow along.
2) Thanks! I was wondering how I would demonstrate the testing regime. Might've been better if I had animated the lines to move with the plane but within my whiteboard skillset I think it was a decent enough compromise
3) Yep, that's the next step. I still need to write up a 2d comparison plotter, which is gonna be annoying lol. I genuinely spent more time writing up the plotters than doing the actual math and fitting cuz idk how to use matplotlib and plotly very effectively.
4) Yeah I went precise nosevelocity --AoA,AoS--> precise IAS --densityratio--> precise TAS. It was 6 months ago when I was takling to AVB on his discord that i realized WTRTI has too many problems and subjective choices made in terms of how its calculations are done that I decided to just stick with the localhost numbers and work from there myself. Very useful logging tool but lags the crap out of my computer and I can't trust the calculated numbers it spits out.
5) It might, but it also might not. In the previous two videos you can see that the old EM Diagram had contours that weren't perpendicular to the x axis. Obviously this is moreso a problem with the lack of accurate and precise data in the midsections but I also don't want to spend hours filling in the empty space in the middle. For the purposes of Air RB, the most important sections are the not-turning and fully-turning anyways, the middle is just sorta there as filler. Of course it would be nice if it were 100% accurate but because the method is inherently messy, I was always just kinda looking for a "close enough". For a more accurate picture, you'd probably need to just start from the other end and solve flight models analytically. I used scipy.optimize.curve_fit, and I'm not really that familiar with it, but when eyeballing the 3D surface plot with the scatter data it looked like a sufficient enough fit.
6) Probably... but I'm also not really that comfortable with sim controls, and I wanted the data to be well-fit for air RB. You can see the scatter data is mostly clustered around the not-turning and fully-turning lines so the fit works best there, which is what I want when comparing air rb performance. If I were using sim joystick controls, I would have to consider things like making sure not to stall, and I didn't want to deal with that. Also I think the curve fit surface came out pretty nicely considering everything so I'm not too fussed about a bit of rolling and yawing. keyboard elevator seemed to work pretty okay.
If anyone were to check out the code then it'd be you lol, its uploaded in the gdrive that I've shared in the description so you could check it out if you wanna
I can see why it's powerful, but i am a plotly noob. maybe if i watched some dedicated tutorials on it but im only using it cuz 3d matplotlib sucks booty
Yea, the vid is good! It might be too advanced if one won't search things up and rewatch, but that's not always a bad thing. If you'll need help with plotly I'll gladly help, tho chatgpt is likely better.
yeah chatgpt is usually okay at pointing me in the right direction tho it screws stuff up too often to be used directly but it tells me what libraries are worth looking into the documentation for
Great work! I don’t have any real complaints today really good job. I definitely wouldn’t have expected that parametric fit to work so well, but honestly it looks pretty great! You could always look at trying some simple interpolation or knn algorithms.
A final thought about overcoming some of these limitations would be to try and build a model using test data in order to implement the analytical method. Its likely gaijin used reasonably simple models in thier drag calculations so that you can probably get a reasonable drag Polar without too much hassle by varying aoa at various speeds and comparing to simple models. For what it’s worth if you can get a good model of drag, the engine data is pretty accessible in the dataleaks, a couple years ago i put together a tool for parsing it. Id also be happy to pass along some info on drag modeling from my aircraft design courses
Thanks! I was considering using a hybrid testing/analytical method. I'm fairly confident I know how drag is calculated atleast for singular parts of the aircraft in gaijin's war thunder. I might pursue it in the future but for now I think I'll stick with this until it becomes too unweildy/messy
Polynomial surface fits work very well on smooth featureless surfaces like SEP in this case. It might be reasonably accurate even with 4th order fit
@@CatWerfer as long as you know the functional form you shouldn’t need to worry about any part buildup, you can just approximate the whole-aircraft drag Polar and use it to recreate all of the different EM charts, thats how they originally did it aswell was just taking drag polars from the engineering groups and validating with flight tests. I think your approach would be very similar to what you’ve just done. This is perfectly fine for props itll only matter with supersonics due to the limitations you already laid out. Either way excited to see your progress!
@@KekusMagnus frequently I prefer to have the functional form motivated by some theory, but can’t argue with results!
4th order didn't work as well, was a bit too smooth and there were regions of visible over and undershoot. 6th is much better and there are no visibly eyeball-able deviations from the scatter data.
Yeah I'll try to think about how I'd approach a hybrid experimental/analytical method for the future but for now I think I'll wrap up this series with this. I was planning on it only being a 2 parter but this has gotten much longer than i thought it would be
For more accurate testing without instructor spikes i would recommend setting up mouse joystick with autotrim, set joystick to square change from simple to standard to allow full motion and remove deadzone and make it not control rudder. Instructor kills turning on some planes and will make results look bad when flight model is different.
Someone else suggested this too but for what it's worth it don't think I care enough to fill out the middle accurately to do sim controls cuz at that point I have to add in a whole other set of restrictions like not going over stall aoa and I'm only really interested in edge behavior for air rb. That probably makes it so these graphs apply less in sim but if the performance is gimped in air rb then I want that to be reflected
Regarding the testing methods: it might be better to use full-real controls/mouse-as-joystick as you won’t have to fight the virtual instructor when you want to find data with only one rotational axis of input. Even with keyboard-only inputs, the virtual instructor still sometimes limits the maximum deflection of control surfaces depending on the air speed and AOA (One example being recovering from a stall with only keyboard inputs)
@lukeskywalket2894 I'm an air rb players so the instructor limits are important to me. Just because the 109 can pull better in sim doesnt mean than it does in rb, and i care mich more about rb performance than for sun. Using sim controls would also mean I'd have to worry about not stalling in flight and I'm also not used to the attached camera view that sim controls puts you in.
If you want to test for sim you can feel free to use my code which is in the shared drive
@@CatWerferah, that makes sense. I’d still recommend using one of the simplified mouse-as-joystick control modes though, since there are ways to get more fine-tuned inputs while still having the limitations of the virtual instructor
@@lukeskywalket2894 maybe ill try it later but not atm
1:31 holy shit
i dont get rocket kills that often but if i want them, the a7m2 is definitely the plane to do it in
Neat
Great, now I can prove why La-7 is superior to Yak-3 using math, thanks man!
Ok I’m interested tell me more please hahahah
This method of calculating SEP should be the default in WRTI
I suggest you try the F8F-1 next for comparison, there is a lot of talk about its flightmodel recently with the drop to 4.7 It bleeds lots of energy with the slightest aoa
Next step is to use these to verify if the flight models are accurate, though gaijin will drag their feet no matter what
The reason they filter and smooth data for wtrti is because its primarily meant to be an addiitoonal hud element and it would be annoying to see a number jittering all over the place. Most people arent gonna care about infinite precision for sep and those who are can prob calculate it themselves, though it would be much nicer if they just made an unfiltered option available. I find the precsision cutoff to be more annoying tbh.
Tbh I don't care too much about realistic flight models too much as long as they don't feel too horrendous to fly. I'm also not really planning on using this to crosscheck flight models since this only reflects air rb performance anyways
@@CatWerfer It's already very jittery and hard to read, especially during climbs unless you decrease the polling rate to 1000ms. I suspect this is due to the low precision of using TAS.
I suspect many meta-defining and/or uncompetitive FMs are the way they are because they are badly modeled (think P-39, 152H, A6M5, or F8F, F6F, 190A5) Gaijin has had plenty of turnover in its FM team in the past decade, FM quality standards are likley all over the place. Such a critical aspect of the game should be taken more seriously
@KekusMagnus i agree, I've suggested changes to the developer again but he seems like he's got more important things om his plate right now.
Yeah flight models are off and it kinda sucks that gaijin is so horrible at balancing but if gaijin dismisses actuakly valid evidence I don't think this will change much until we can get more complete and accurate data
Yeaah let's go
Is it possible to isolate the radial component of acceleration, for the turnrate calculation?
yeah it is possible since you can log pitch and roll angle of the aircraft as well, it would probably give a more accurate turnrate. However for the purposes of comparing aircraft i think what i have here is probably alright
Do you know the TH-cam channel "Gregs airplanes and automobiles" ? Its awesome, its like a university class about aviation, focusing in WW2. It studies in great detail the performance of WW2 aircraft including EEM charts, engine technology performance evolution, human factors in cockpit design, etc. Greg has like 8h of content just about the p47 alone. His content leaves me with my jaw dropped. If you made this video you will like his channel for sure!
Yeah I saw some of their stuff, really high quality! I might try something like that if I have the mental fortitude to spend 8hours talkinga bout one plane lmao
Let him COOK
with the temps my laptop was running at i could probably fry an egg for ya