Scared of Turbulence? AI -Might- Fix That

In this episode, we explore how AI can help aerial vehicles adjust to extreme turbulence in real-time and enhance safety/stability, even in challenging driving conditions.

Episode Notes

(0:50) - AI Beats Turbulence

Become a founding reader of our newsletter: http://read.thenextbyte.com/

Transcript

What's up friends? Are you scared of turbulence while flying? You're not alone. Don't worry though, because we're talking about a team today from Caltech and Nvidia that are working on an algorithm called FALCON that's gonna fix it and make it a smooth flight for you even when there's bumpy air. So, fasten your seat belts, stow your tray tables and move your seat to the upright position because we're gonna fly into this one.

I'm Daniel, and I'm Farbod. And this is the NextByte Podcast. Every week, we explore interesting and impactful tech and engineering content from Wevolver.com and deliver it to you in bite sized episodes that are easy to understand, regardless of your background. 

Daniel: What's up folks, how's it going? Today we're talking just like we said, all about turbulence and flying and airplanes and technology that's gonna fix it. And this hits home for me. I actually just got off a plane like two hours ago and we hit a little bit of rough air but there's always that person on the plane. Fortunately, it's not me, but there's always that person on the plane sitting near you who you can tell is like deathly afraid of turbulence. And they think, this bumpy air, the plane's about to fall out of the sky. And in some ways, turbulence as like something scary has been massively overblown, but it's also still a pretty big safety risk for aircraft. So, it's not likely that the aircraft is going to fall out of the sky and crash because of turbulence, but it's actually quite frequent that folks get an injury in some sort, or at least an inconvenience. They spill their drink on them because of turbulence in an airplane. I think it was a Singapore Airlines flight recently where like a hundred of the passengers were hurt because of turbulence. That was pretty insane. Turbulence happens really quickly, pretty unpredictable. There's lots of things that can cause it. And with climate change, actually these micro changes in airflows are meeting, these events are happening quite more frequently than in the past. So, turbulence is getting worse. It's an annoyance. And it's also a safety risk for passengers on a plane. So, engineers want passenger planes to be able to handle this, but also think about in terms of like drones flying around that, you know, if you've ever tried to fly a quadcopter drone in the wind, it's actually pretty scary to be like, oh, I spent 500 bucks on that thing. And it's bobbing around in the air because of the wind. We’re working about it, talking about a team of engineers that want to be able to help these aircraft sense and adjust to turbulence. Similar to how birds do in the sky to fly smoothly and improve safety.

Farbod: Yeah. Yeah. Real quick, when you're talking about the person next to you that's deathly afraid of turbulence, that was me for the longest time. As someone being from Iran and I would go back every summer, 15-hour flights, miserable dude, would literally not sleep a blink of an eye. I was like holding on to the seat for dear life because every shake, like now I know it's not that bad, but like you're what, thousands of feet up in the sky, things are shaking, you have no control, it's scary. And it's, I totally get why people feel like.

Daniel: I'm telling you, Nellie and I, we were recently in Alaska and we took a flight on a bush plane, like a little pond hopper plane from, where was it? Just outside of Homer, Alaska into Katmai National Park. And it's like seven passengers in the plane, two or three rows of seating, like really, really small prop plane. And I'll tell you when we're only like 300 feet above the ground and you catch bumps in the air, you're like, oh man, this is really scary. And there's actually this guy in front of us, in front of Nellie and I, who had his finger hooked into a D-ring on the side of the wall. I'm like, dude, if we hit a bump in the air, all you're going to do is deglove your finger.

Farbod: Oh man. Yeah.

Daniel: But all that being said, right, there's lots of different aircrafts that can benefit. If there were some way for them to be able to sense and adjust the turbulence. There's actually, from a technical perspective, it's quite challenging. They've tried this before with different AI methods and different sensing, and essentially, they have only ever been able to react to turbulence by feeling it, right? So, by seeing it, detecting it with sensors, and because turbulence happens in so many different scenarios with different causes and different changes in airflow, essentially, these old AI models would need to react to what they see. They're reacting too slow and they needed a ton of training because there were so many different situations about this type of turbulence or that type of turbulence. Essentially the reason you experience turbulence in bumpy air is because of unpredictable changes in the air current around the plane. By nature of them being unpredictable, they're really, really hard to empirically train an AI model to say, oh, these are all the different ways that turbulence can happen because there are millions of different possibilities, which is why they've had challenge so far with AI methods to try and create control algorithms for planes to fight against bumps in the air.

Farbod: Yeah. I mean, you know, you have laminar flow, which is nice, steady, smooth, exactly what you want to see going back to fluid mechanics. Then you have the turbulent flow. Like you have weird pockets of air out there, basically speed bumps for the airways. That's what you're running into. And the problem, at least in my mind, breaks down into two portions, right? You have the ability to detect that turbulent flow is coming your way and the patterns in which it's coming, then you have the response to it. You were talking about earlier on like nature and how that we have birds that can react to turbulent flow, right? Like how do they do that? They have wings and those wings can be extended or brought in to counteract some of that turbulent behavior. And that's kind of what this team is going for, right? Like if we can start delving into their sauce, you already mentioned, empirically determining it's difficult, there's two bits to the sauce. You have the detection and then you have the actual control to counter it.

Daniel: And, and one thing I want to mention just back to like it being more than just airplanes, passenger aircraft that can benefit from this. As they mentioned, like if you're flying UAVs around tall buildings that wind changing in cities and like narrow airways between two tall buildings, just trying to contextualize how complex the turbulence problem can be when you brought in the scope to include UAVs as well, not just passenger aircraft. When you physically have hard obstacles that are redirecting wind around it that are included in this fluid dynamics question, it becomes a lot, lot, lot more challenging, which is why they can't determine it empirically. They have to use this new method, which uses Fourier methods to try and break down wind patterns into waves, which is waves are things that AI model, you know, it's math, it's on a chart. It's way easier for a math model and AI model to understand waves. And that's how it's doing the detection because otherwise if you're just trying to use wind speed and wind direction as your vectors to understand and train this thing, especially when you're talking about UAVs and cities, right? You're gonna have way too many different factors that you're never gonna be able to understand it.

Farbod: Definitely. Also, I don't think there's anything more triggering for an engineering student than Fourier, the Fourier transform.

Daniel: You said Fourier transform. That's something that did hurt me a little bit. But honestly, when you said fluid dynamics, that hurt me a little more. Just the whole subject in general, it's the worst grade I got in all of school ever.

Farbod: That's a testament to what I just said.

Daniel: I'm uniquely unqualified to speak about this.

Farbod: But the feedback from the team in terms of how they use it, that was really impressive to me. Basically, treating turbulence as kind of like a sinusoidal wave. And then that's so much easier to analyze. Like even if you're just visually looking at it, you're like, all right, this goes up for this period of time, then down for this period of time. And then if you now have a mechanism that can counteract that movement, boom, you have stability or relative stability. So that is in itself pretty genius. That I wanted to highlight. You were talking about inside cities and quadcopters and stuff like that, why turbulence is important. If you have a massive vertical obstacle against wind that's coming in at very, very high speeds, that is like the perfect recipe for turbulence, right? Because you got it hitting the wall, bouncing everywhere. And I don't, at first it didn't make sense to me why they were so fixated on quadcopters, but now thinking about like Amazon drone deliveries and stuff like that, that makes a lot sense.

Daniel: And everyone says the future of transportation is VTOLs, vertical takeoff and landing passenger craft.

Farbod: Yeah, good point.

Daniel: Like, hello, you're starting to take a massive portion of the traffic that's on the road and put them in the sky and cities are uniquely where there's a ton of population density and where people would want to use these things anyway. Like I imagine tons and tons of challenges, challenges that exist with helicopters and that's why helicopters have a really high crash rate. So definitely easy to contextualize what the, so what is here on this one.

Farbod: For sure, for sure. And like, if you think the speed bumps in your neighborhood are bad, just wait until you hit a turbulent flow when you're on your flying car on the way home and potentially it can be life threatening. That'll really put a bump in your step.

Daniel: Well, and so to go back to their secret sauce here, we mentioned Fourier methods and it's interesting like they were essentially able to break down different patterns in the wind into waves. They modeled how changes in wind happen, basically the changes in the energy in the wind by modeling the frequency of the wind. They noticed that turbulence is typically related with low frequency changes in the energy. One of the things that was really interesting to me is this approach was not focused on previous methods are trying to empirically understand, like, oh, these are the different types of turbulence. And this is how you react when you see this certain type of pattern. In this case, their method, which they call FALCON, I forget what it stands for.

Farbod: I got you Fourier Adaptive Learning and CONtrol.

Daniel: Pretty, pretty awesome. If you remember that their inspiration is trying to help UAVs and passenger aircraft and commercial aircraft adjust to win just the way that birds do. So, choosing a bird name as the name for your secret sauce.

Farbod: I can give it a nine. I give it a nine. That's solid.

Daniel: 9.2 from my end, incredible. But FALCON is basically trying instead of to categorize all the different types of turbulence and then basically create a playbook for aircraft and how to react to different categories of turbulence. Instead, they said, let's just understand the fundamentals of the energy shifting, what's going on in turbulence. So instead of, in my mind, instead of creating like a dictionary of all the different words, right? You're trying to fundamentally teach the AI model how the language works or what the language is. And that's what they've done here and with some success, right? So, let's talk, I guess, did you have something else in the secret sauce? Right? You mentioned sensing and control, right?

Farbod: I was just going to say the control bit, but I think in the so what, you're going to mention that control bit.

Daniel: Yeah, well, exactly. I wanted to go to testing, right? So, you have this “So what you're able to plot turbulence as waves?” How does that impact your ability to react to actual turbulence in a testing scenario, in a physical testing scenario? So, what they did is they tested this algorithm FALCON in a wind tunnel and I think we forgot to mention that it's Caltech and Nvidia researchers working together on this right?

Farbod: Yes.

Daniel: Given how excited we got no kudos to the to the author so far so I will distinctly give a kudos to the team from Caltech. They used an air tunnel, or wind tunnel, and they used an air foil, so basically just a physical model of a wing, a cross section of a wing, equipped with a ton of different sensors, as well as control surfaces. So, they've got a bunch of sensors on it that are able to inform the model on how the energy is shifting, how the wind is shifting around it, and then they've also got a bunch of control surfaces, basically controllable with motors to physically change the shape of the wing to try and react to the turbulence.

Farbod: Like as the average listener, like you've seen control surfaces when your plane is taking off and landing, right?

Daniel: And I don't know if this is true, right? Cause you and I are engineers and we always try and choose a seat with a view of the wing.

Farbod: Yeah, good point. Actually, good point, good point. But if you've ever seen the wing extend and then when you're landing, the flaps come up, right? Like, control surfaces.

Daniel: Yeah, exactly. So, if you've never paid attention to the wing on a plane, I urge you, next time you book a plane seat and you have seat selection, I urge you to choose a seat near a window with a view of the wing. Because it's actually quite fascinating to watch these control surfaces, like you're saying. They've got different flaps. They've got different rudders. They've got different types of adjustment to try and physically change the shape of the wing. And that's all they need to control the plane. Which is just crazy, blows your mind that they can move such a big, large, heavy object just by changing the shape of the wing. But that's what they do, right? And so, they have these control surfaces motorized that they're able to change the shape of the wing in response to the turbulence. They created an unpredictable turbulent environment by essentially placing, I think it was a cylinder with a movable attachment inside the wing tunnel. So, it's like flapping around, moving around, creating unpredictable wind that gets all twisted up. And when we say turbulent wind, it's really just like the air flow, instead of moving in clean straight lines, laminar flow, like Farbod mentioned earlier, you've got all these like weird squiggles and mess, like imagine if you took a ball of yarn on the table, and you're used to the yarn being in like long straight parallel lines, think like the strings on a guitar and that's laminar flow. Turbulent flow is like someone went and scrambled all those up and then this wind is hitting the wing but instead of it being nice and smooth and easy, it's all scrambled around. And that's what causes the plane to bump around. So, they're creating turbulent wind in the same way by having the cylinder that moves around with this attachment in the wind tunnel, unpredictable wind. And they essentially turn this FALCON algorithm on. Daniel: And they said, like, understand how this turbulence is working. And then learn how to control and stabilize the UAV. I think even in these, like, super extreme conditions, it was able to control and stabilize the UAV in under 10 minutes, I think is what they said.

Farbod: That's correct.

Daniel: I have no context as to how long it takes other models to do this, but the fact that it's completely random, completely random turbulence, and then FALCON is able to actively stabilize the UAV after only nine minutes of training time is insane to me because the other way of like empirically trying to understand all the different types of turbulences, I can only imagine the thousands and thousands of hours of testing and flying that you would have to do to create all the different scenarios, right? A dictionary of all the different scenarios that could potentially happen to train the old AI method.

Farbod: Correct. And I think it's worth noting, at least if I remember correctly, this isn't like a binary approach either. It's not like for the full nine minutes, FALCON is bad at handling turbulence. And then suddenly it becomes good and all is stable. It's progressively getting better as more and more data is coming in about the turbulent flow and it's better processing it. And as it's adjusting to the flow, it's getting more feedback, so on and so forth. So, you have this curve that's hockey sticking in your favor for stability. Nine minutes, I don't know, like you were saying, it seems pretty impressive for me as is, but the researchers did mention that, that is the parameter that they wanna keep optimizing over time, because when a turbulent event happens, like the one with Singapore Airlines, you don't get the opportunity to gather nine, 10 minutes of data. You want to react as quickly as possible.

Daniel: Yeah, and I wonder if it's like you have nine minutes of training for this type of wing geometry and this type of wind speed. Instead of trying to categorically and empirically look at all the different types of potential turbulence and all the different settings, in this case maybe you're just looking at a few key variables that you can pre-train a model and deploy it into the field on a bunch of different platforms, let's say, and reduce that training time. I hope that this is the case, right? That if they were to take this same airfoil with the same geometry in a similar situation in a different wind tunnel, maybe the training time would only be one minute instead of nine.

Farbod: That's my hope as well. It kind of reminds me of an article we had, I think, two, three months ago at this point, where there was this generalized robot that came pre-programmed with some basic instructions on how to do a task. And then in situ, it was able to fill the remaining 25, 30% of its training instead of learning from the ground up. Like you said, I'm hoping that whatever model they make is able to have some level of understanding based on the airfoil, weather conditions, whatever, on how it should be trying to counter turbulence, but then it picks up the remaining on the way. But this, again, as far as I understand the aerodynamics, there's so many factors that go into this that I wonder how much can they really meaningfully pre-train before deployment?

Daniel: Yeah, and I mean, I'm assuming that there are a number of different steps they can take here to make it applicable to commercial aircraft, like you're saying. You don't want after, you don't have to wait nine minutes in a commercial aircraft dealing with severe turbulence before you can understand what's going on, right? At that point, the pilot's probably just gonna change the altitude to try and get out of the turbulent streams to save the passengers from getting scrambled around in the back. But I think even in the commercial aircraft, in the UAV space, like if you have to fly a drone between two buildings to do some type of thermal assessment, right, you're not able to easily change the altitude or just go to a different spot. Like you have to fly the drone in that spot, even nine minutes of training time would be, I would say, not an insurmountable task. It would be surmountable, right? If you're able to control the drone for nine minutes and not crash into two buildings for nine minutes, then it can truly adapt and understand the turbulent flow and be able to control and stabilize itself. Like there, maybe in the commercial aircraft space, there's more work that needs to be done, but I bet in terms of testing with quadcopter UAVs, they can probably deploy this today and start to see an immediate impact.

Farbod: You're definitely right, yeah. And I'm excited to see it as soon as possible honestly. Like I said, used to be afraid of turbulence, gotten better with it, still makes me uncomfortable at times. And I for sure don't want to be in the same position as those 100 people that got injured on the Singapore Airlines flight.

Daniel: Well, and to me, the analogy, it kind of reminds me of something similar that I've seen in the automotive space. Have you seen these cars that, now that we've got vehicles equipped with so many sensors and cameras, they can kind of detect where there are bumps in the road and then automatically adjust like the ride stiffness and ride height to make a smoother ride. Have you seen any of these demonstrations at all?

Farbod: No, no, I haven't.

Daniel: It's pretty interesting. But to me, it feels really similar to this where it starts as like a cool demo that someone does saying like, oh, in a few years, we're gonna roll this out. And now this is a feature that I see a lot of OEMs working on and starting to roll out into their vehicles, like starting last year, I think, where they've got vehicles that have like active ride stabilization, trying to look at obstacles, bumps in the road, potholes as an example, and then can like automatically adjust the ride stiffness for that tire that's gonna hit the pothole to make sure that your car doesn't go to dunk when it hits it. Which is-…

Farbod: Don't tell me that, man. Don't tell me that. I got my rose-colored glasses on for my 2013 Corolla. You're ruining this for me.

Daniel: Well, you're gonna keep hitting potholes in that thing for 300,000 more miles before it dies. In my mind that's a that's more of like a I'm not going to say one dimensional problem. A two dimensional or three-dimensional problem. Whereas the Aerodynamics problem is a little bit more complex. But if that has already been figured out and started to roll out on passenger vehicles on the road, I could see this being something very similar that follows in the sky.

Farbod: No, that's a that's a really good point. What do you say we start wrapping up?

Daniel: Yes, sir. Let me wrap us up here and just give everyone a reminder, if you are scared of sudden turbulence in flight, you're not alone. Farbod’s in that group of people scared of turbulence. And recently over a hundred passengers were hurt on a Singapore Airlines flight. I think it was in May, just because of turbulence. So, turbulence can hit without warning. It almost always does, especially in rough weather. And it's actually getting worse with climate change. So, engineers are working on planes that can actually understand the dynamics behind the turbulence, adapt to it, just like the way that birds do. They call their algorithm FALCON, which is pretty cool. It's like trying to learn from birds. And they say that they can deploy this to future aircraft to adjust their movements, basically change the shape of the wings to make flights smoother and safer using advanced sensors and AI. So, imagine flights that are like as calm and smooth as a car ride in a Rolls-Royce thanks to this innovation called FALCON. So flying is gonna get safer, the future looks bright. I don't think you should be as scared of turbulence in the future if this plane has got the FALCON algorithm.

Farbod: Money! I want to play with FALCON on it, take the MCAS out of the Boeing 737s and put Falcon on it.

Daniel: No comment.

Farbod: Alright folks, thank you so much for listening and as always, we'll catch you the next one.

Daniel: Peace.

As always, you can find these and other interesting & impactful engineering articles on Wevolver.com.

To learn more about this show, please visit our shows page. By following the page, you will get automatic updates by email when a new show is published. Be sure to give us a follow and review on Apple podcasts, Spotify, and most of your favorite podcast platforms!

--

The Next Byte: We're two engineers on a mission to simplify complex science & technology, making it easy to understand. In each episode of our show, we dive into world-changing tech (such as AI, robotics, 3D printing, IoT, & much more), all while keeping it entertaining & engaging along the way.