Latest / Mechanical Engineering Made Simple / (#101) Fixing Teslas and Evolving NASA Parts

Transcript

(#101) Fixing Teslas and Evolving NASA Parts · Mechanical Engineering Made Simple

• 0:00
  Welcome back design junkies, to Mechanical Engineering Made
• 0:02
  Simple. If you enjoy the podcast, show
• 0:05
  your support with a like a follow or comment.
• 0:07
  It's free for you but means the world to us.
• 0:10
  Let's keep leveling up together. It is great to be back.
• 0:12
  So today we are looking at a well, a massive shift in how
• 0:17
  things are made. And look, I know we hear game
• 0:19
  changers every other week in this industry all the time.
• 0:22
  But today, I don't mean a small change or, you know, a new
• 0:25
  button in SolidWorks. I mean a fundamental rewrite of
• 0:29
  the rulebook. Yeah, total paradigm shift.
• 0:31
  Exactly. We aren't just talking about CAD
• 0:33
  drawings anymore. We're talking about systems that
• 0:36
  evolve geometry the way nature evolves bones, right?
• 0:39
  Cars that can physically fix themselves while sitting in your
• 0:43
  driveway and the AI Co pilots making it all happen.
• 0:46
  That's right. And our mission for this deep
• 0:48
  dive is really to walk through that story, the story of the
• 0:51
  modern digital engineer. OK, we're going to trace this
• 0:54
  chronologically, so it all makes sense.
• 0:56
  We'll start with the birth of the digital twin at NASA.
• 0:59
  Back in the space race. Back in the space race, then
• 1:02
  we'll move into generative design where the AI actually
• 1:06
  takes the wheel. And finally we have to do this.
• 1:08
  We look at the risks. The scary stuff.
• 1:11
  The scary stuff What happens when these incredibly powerful
• 1:14
  tools get into the wrong hands? I was reading through the source
• 1:18
  material you sent and I have to say there were some serious aha
• 1:22
  moments. Specifically, a NASA bracket
• 1:25
  that looks like it belongs in an alien autopsy and a ransomware
• 1:29
  concept that learns how to play you like a video game.
• 1:32
  But OK, before we get there, let's start with the Foundation,
• 1:35
  the Digital Twin. The digital twin.
• 1:37
  It's a buzzword we hear constantly.
• 1:39
  Right. And usually when I hear Digital
• 1:41
  Twin, I just think of a really high fidelity 3D model, you
• 1:44
  know, a fancy assembly and KTS spinning on a screen.
• 1:48
  But you're telling me that's too shallow?
• 1:50
  Way too shallow. If you just have a 3D model, you
• 1:52
  have a digital shadow. Not a twin, a shadow.
• 1:54
  OK, a true digital twin or DT, is a synchronized link with a
• 1:59
  physical object. Imagine a turbine in a factory
• 2:03
  or a car on the road. The DT uses data from embedded
• 2:06
  sensors to simulate the the real time status of that object.
• 2:11
  So it's not just the geometry, it's the state.
• 2:14
  Exactly. If the physical object heats up,
• 2:16
  the twin heats up. In the simulation, if the
• 2:18
  physical object starts vibrating at, say, 40 Hertz, the twin
• 2:22
  vibrates at 40 Hertz. It's a living mirror.
• 2:25
  That distinction is crucial. That is the difference between a
• 2:27
  photograph and a live video feed.
• 2:29
  Perfect analogy. But where did this start?
• 2:31
  Because I'm guessing we didn't have IoT sensors back in the
• 2:34
  day. We did not, but we did have the
• 2:36
  Apollo project, and this is the origin story.
• 2:39
  OK, NASA had this massive logistical problem.
• 2:42
  I mean, once you launch a vehicle into space you can't
• 2:45
  exactly pop the hood to check on a rattling noise.
• 2:48
  No physical access. Zero.
• 2:50
  So essentially built two of them, two space vehicles, one
• 2:53
  went on the mission and one stayed right here on Earth.
• 2:56
  The twin. That sounds incredibly
• 2:58
  expensive, but I guess it was necessary.
• 3:00
  So if an oxygen tank explodes on Apollo 13.
• 3:03
  You test the fix on the Earth twin first.
• 3:06
  You see if it works before you tell the astronauts 1,000,000
• 3:09
  miles away to try it. It mirrored operations to allow
• 3:14
  for troubleshooting. That was the grandfather of the
• 3:17
  digital twin. High stakes, high cost, and all
• 3:20
  manually synchronized. Exactly.
• 3:22
  OK, that's NASA. They have the budget for two
• 3:23
  spaceships. But how does this Fast forward
• 3:27
  to today? How does this concept look in my
• 3:30
  driveway because I don't have a second car in a lab mirroring my
• 3:34
  commute? No but a server does and the
• 3:36
  best examle right now is Tesla secifically the model right?
• 3:39
  Tesla uses digital twins to design their cars but the magic
• 3:43
  haens because they Kee them connected.
• 3:45
  The cars are constantly talking to a remote center and this is
• 3:48
  more than just GPS or Spotify updates.
• 3:51
  Much, much further. So here's a specific story that
• 3:53
  really highlights the power of this Back in 2013, there were
• 3:56
  reports of three Model S cars experiencing burnouts.
• 4:00
  Burnouts. So fires.
• 4:01
  Essentially overheating issues that led to fires.
• 4:03
  Yes, yes. Now in the old world of
• 4:04
  automotive engineering, what happens next?
• 4:06
  That is a massive physical recall.
• 4:09
  You're sending letters. People are driving the
• 4:11
  dealerships waiting for hours. It's the logistics nightmare.
• 4:14
  And APR disaster. But Tesla used the digital twin
• 4:18
  approach. They analyzed the data streaming
• 4:21
  from the cars, the digital twins of the entire fleet.
• 4:25
  So what did they find? They realized the problem wasn't
• 4:28
  a defective part necessarily. It was an issue they could fix
• 4:31
  by changing the suspension clearance at highway speeds and
• 4:36
  altering the charging settings. So the car was riding too low
• 4:39
  and hitting debris on the road. Partly, yeah.
• 4:41
  So they pushed an update over the air.
• 4:43
  They effectively fixed a hardware vulnerability with
• 4:46
  software. The problem was solved without
• 4:48
  the owners ever having to open the hood.
• 4:50
  That is wild. I mean, it just changes the
• 4:53
  entire economics of maintenance. It does, and it lets you do what
• 4:56
  we call predictive maintenance. You can re engineer the life of
• 4:59
  a product like doing fatigue analysis on an aircraft wing
• 5:03
  without destroying the physical object.
• 5:04
  You're simulating the stress on the digital version to save the
• 5:07
  physical one. Precisely.
• 5:08
  It's a game changer. It's like having a voodoo doll,
• 5:11
  but you use it to heal the person instead of hurt them.
• 5:14
  That is actually a surprisingly good analogy.
• 5:17
  I might have to steal that one. Feel free.
• 5:20
  OK, so that's monitoring things. That's keeping an eye on what
• 5:23
  already exists. But the sources you sent over
• 5:26
  suggest an even bigger shift is happening in how we create
• 5:29
  things. This brings us to generative
• 5:32
  design. This is where the role of the
• 5:34
  engineer really starts to change.
• 5:36
  Traditionally, engineering is well, a human sits down, stares
• 5:40
  at a blank screen and draws a shape.
• 5:42
  I think this bracket should be a square with a hole in it.
• 5:45
  Exactly. Then they run a simulation to
• 5:47
  see if it breaks. Right iteration one it breaks
• 5:49
  you make it thicker iteration 2. It's manual trial and error in
• 5:53
  generative design. The human doesn't draw the shape
• 5:55
  at all. So what do they do?
• 5:56
  Just ask the computer nicely? Yeah, please make a bracket.
• 5:59
  Almost. They define the requirements.
• 6:01
  They tell the computer I need a part that can hold this much
• 6:04
  load, uses this material, and very importantly stays out of
• 6:08
  these specific zones. And then the AI generates the
• 6:13
  shape. It sounds a bit like those AI
• 6:15
  art generators everyone's playing with now.
• 6:17
  Is the exact same concept. One of our sources uses the
• 6:20
  analogy of Dolly. You type in an armchair in the
• 6:24
  shape of an avocado, and it creates it.
• 6:26
  It explores the latent space of possible images to find one that
• 6:31
  matches your text. An avocado armchair.
• 6:34
  I would sit in that. Sounds comfortable, but in
• 6:36
  engineering, we're not making avocado chairs, we're making
• 6:39
  titanium space parts. The stakes are a lot higher
• 6:41
  because physics doesn't negotiate.
• 6:43
  Good point. So give us a walkthrough of a
• 6:45
  success story. I'm a little skeptical.
• 6:47
  Is the AI actually finding better solutions or just weird
• 6:51
  ones? Let's look at the NASA Goddard
• 6:53
  Space Flight Center. They have this process called
• 6:56
  evolve structures. Okay, they were building
• 6:58
  something called the Excite tip tilt bracket for a balloon
• 7:01
  mission The. Excite Bracket sounds intense.
• 7:04
  What were the constraints? It had to hold a sensor short,
• 7:06
  but the main challenge was shock.
• 7:08
  It had to survive a parachute landing.
• 7:10
  We're talking about A10G vertical load. 10 GS.
• 7:13
  So if that bracket weighs a pound, it suddenly feels like 10
• 7:16
  lbs instantly. Right.
• 7:18
  And it couldn't block the light paths for the instrument.
• 7:22
  So NASA set up a showdown Man versus machine.
• 7:26
  I love a good rivalry O the human team.
• 7:29
  An expert designer and a structural analyst.
• 7:32
  People at the top of their game. And the AI team.
• 7:34
  One engineer and his only job was to encode the requirements
• 7:38
  into the software. OK, So what happened?
• 7:40
  The human team worked for two days.
• 7:42
  They went through four different iterations.
• 7:44
  They struggled to balance the stiffness and the weight.
• 7:46
  And AI team it. Took about 30 minutes of human
• 7:49
  time to set up and then the AI just ran for an hour.
• 7:52
  So it's faster. I mean, that's not surprising,
• 7:55
  but was the design actually better?
• 7:57
  Significantly better, right? The human designs were either
• 8:00
  too heavy or in their final attempt, just unmanufacturable.
• 8:04
  It looked good in CAD, but you couldn't actually build it.
• 8:07
  The AI design however, well, it looked weird.
• 8:10
  It looked organic. It looked bone like.
• 8:13
  When you say bone like what do you mean?
• 8:14
  It looked like the inside of a femur, all these messy,
• 8:16
  intricate structures. And that happens because the
• 8:19
  algorithms used topology optimization.
• 8:22
  They just remove material that isn't carrying stress, which is
• 8:25
  exactly how nature evolves bones.
• 8:28
  It's a principle called Wolf's law.
• 8:30
  If a part of your bone isn't bearing a load, your body
• 8:32
  dissolves it to save energy. And if it's under high stress,
• 8:36
  it adds more material there. Exactly.
• 8:39
  So the AI is basically simulating millions of years of
• 8:42
  evolution in an hour. Doesn't care about looking
• 8:45
  pretty or geometric, it only cares about the load path.
• 8:49
  And what were the final stats on that?
• 8:50
  The AI part was three times stiffer and significantly
• 8:54
  lighter than the best human attempt.
• 8:56
  Plus it was designed for digital manufacturing, so they can send
• 8:59
  the file straight to a 3D printer.
• 9:01
  That's incredible. You're talking about development
• 9:03
  time dropping from months down to days, and you get a better
• 9:06
  part. It's not just about speed, it's
• 9:08
  about breaking out of human bias.
• 9:10
  Yeah, we like straight lines and 90° angles.
• 9:13
  Nature doesn't. But this raises a tough
• 9:15
  question. If the AI is so good, faster,
• 9:18
  stronger, lighter, why do we need the humans at all?
• 9:22
  Why not just hit print on a car? Because of something called the
• 9:25
  complexity trap. OK, let's unpack that.
• 9:27
  What's the complexity trap? Think about a modern car.
• 9:29
  It's not just an engine of wheels anymore.
• 9:31
  It has 80s advanced driver assistance systems.
• 9:34
  It has heated seats, infotainment, hundreds of
• 9:37
  sensors. All of that needs power and
• 9:40
  data. Which means wires.
• 9:41
  Miles of wires. The Siemens source highlights
• 9:44
  this perfectly. The wiring harnesses in modern
• 9:47
  vehicles are massive. They're incredibly intricate.
• 9:50
  You simply cannot draw every single wire manually anymore.
• 9:54
  So you reach a point where the human brain is the bottleneck.
• 9:58
  It is the bottlenecks. You need something called wiring
• 10:00
  synthesis. You set automated rules that
• 10:02
  generate thousands of connections and splices.
• 10:04
  But here's the catch. And this brings us to this idea
• 10:07
  of next Gen. AI.
• 10:09
  And how is next Gen. AI different from say, the chat
• 10:12
  bot I used to write emails? Well, current AI, like those
• 10:15
  chat bots, is based on large language models.
• 10:18
  Yeah, they're great at text and images, but they they struggle
• 10:21
  with physics. They hallucinate.
• 10:23
  What does a hallucination look like in an engineering context?
• 10:26
  It. Looks like the AI is suggesting
• 10:28
  a material that doesn't exist or designing a bolt that goes
• 10:31
  straight through a fuel tank, right?
• 10:32
  Not good. You don't want an AI
• 10:34
  hallucinating the brakes on your car.
• 10:36
  Definitely not. So we need constraints.
• 10:38
  We need first principles. NextGen AI is designed to adhere
• 10:42
  to physics and manufacturing constraints.
• 10:44
  It has to be grounded in reality.
• 10:46
  So where does the engineer fit in this new world if the AI is
• 10:50
  generating the geometry and checking the physics?
• 10:53
  The role shifts. The engineer becomes a copilot.
• 10:56
  They stop doing the rote tasks like verifying every single bolt
• 10:59
  or drawing every wire. Instead, they define the
• 11:02
  problems. They become the architect of
• 11:04
  constraints. Architect of constraints.
• 11:07
  I like that title, sounds important.
• 11:08
  It is. It's a human centric approach
• 11:10
  approach. The AI points the engineer in
• 11:12
  the right direction, but the human has to be there to stop
• 11:15
  the AI from designing something unsafe.
• 11:17
  You need that oversight. So the AI answers the question,
• 11:20
  but the human has to ask the right question.
• 11:22
  That's a perfect way to put. It OK, so the engineer is the
• 11:25
  pilot, AI is the copilot. But looking at the report on the
• 11:29
  malicious use of AI, there's a flip side.
• 11:31
  If this AI is so powerful for building things, surely it's
• 11:35
  just as powerful for breaking them.
• 11:37
  Unfortunately, yes. This is the dark side of digital
• 11:40
  engineering. The same algorithms that
• 11:42
  optimize a bracket for strength can optimize an attack for
• 11:45
  destruction. And what does that look like in
• 11:47
  practice? Well, think about how the AI
• 11:50
  designed that NASA bracket. It explored thousands of options
• 11:53
  to find the optimal shape. Hackers can use that same logic
• 11:57
  for automation of vulnerability discovery.
• 11:59
  Meaning just as the off ago AI found interesting moves in the
• 12:03
  game of Go that no human had ever thought of, right, an AI
• 12:06
  can scour software code to find interesting vulnerabilities that
• 12:10
  no human security analyst ever noticed.
• 12:12
  So it finds the open window that nobody knew was there.
• 12:16
  Then it exploits it. Yeah, the report outlines A
• 12:18
  hypothetical scenario they call Wanna laugh.
• 12:21
  Sound like a bad joke? A very expensive one.
• 12:24
  Imagine a ransomware that uses AI V8 I to constantly rewrite
• 12:27
  its own code. Usually security comanies find a
• 12:31
  pattern, a signature and write a atch to stop malware.
• 12:34
  OK, but if the malware learns and evolves like a video game
• 12:38
  player, it can bypass those patches instantly.
• 12:41
  It evolves faster than the defense.
• 12:42
  It's polymorphic. That is the fear.
• 12:45
  And it's not just software. We're talking about physical
• 12:47
  threats too. Drones.
• 12:48
  I've seen the videos of drone swarms that look like a flock of
• 12:51
  birds. It's terrifying.
• 12:52
  And it's a very hard thing to stop.
• 12:55
  In 2018, there was an assassination attempt on the
• 12:57
  president of Venezuela using drones.
• 13:00
  In 2019, there was that attack on the Saudi Aramco oil
• 13:03
  facilities. And that was a swarm, right?
• 13:05
  Yes, a swarm of 25 drones operating in unison.
• 13:09
  When you combine that with AI navigation, you get a weapon
• 13:12
  that is very cheap but very destructive.
• 13:16
  The swarm just follows simple rules.
• 13:17
  Don't hit your neighbor. Fly toward the target, but the
• 13:20
  group behavior is incredibly hard to defend.
• 13:23
  A guess. And there is something in the
• 13:24
  notes about face morphing that sounded like it was straight out
• 13:27
  of a spy novel. It's a serious identity threat.
• 13:30
  You take two faces, say yours and mine, and you merge them
• 13:33
  into one photo. It looks enough like me to pass
• 13:36
  a human guard, but enough like you to fool the passport control
• 13:40
  AI. So one passport could work for
• 13:41
  two different people. Exactly.
• 13:43
  It tricks the border security AI.
• 13:45
  The take away here is that as we build these digital twins and
• 13:48
  use AI to design our world, we are increasing the attack
• 13:52
  surface. So we need AI fighting AI.
• 13:54
  We do. We need active countermeasures.
• 13:56
  We can't rely on human speed anymore.
• 13:58
  Before we wrap up, I just want to touch on the how.
• 14:01
  We've talked about AI evolving brackets and learning
• 14:03
  vulnerabilities, but strictly speaking, how does a machine
• 14:07
  actually do that? What's the math behind the
• 14:09
  magic? It's actually surprisingly
• 14:12
  intuitive. It's based on trial and error,
• 14:15
  which is a lot like reinforcement learning.
• 14:17
  Think of a SOAP bubble. OK, I'm picturing a bubble.
• 14:19
  Wand, if you dip a wireframe into soapy water, the film
• 14:23
  naturally forms a shape, right? It doesn't know math, it doesn't
• 14:26
  have ACPU, but it naturally finds the shape with the minimum
• 14:29
  surface area. It optimizes itself to the
• 14:32
  lowest energy state. It solves the physics problem
• 14:34
  just by existing, right? Now, in AI, we use a strategy
• 14:38
  that's often compared to a bacterium.
• 14:40
  There's this biological concept called run and twiddle.
• 14:42
  Run and twiddle, yeah. If a bacterium is moving towards
• 14:46
  food, so if life is getting better, it keeps running
• 14:49
  straight. If life gets worse, if the
• 14:51
  chemical signal gets weaker, it stops and twiddles.
• 14:54
  It just moves randomly until it finds a better direction.
• 14:56
  That's. Actually a really good life
• 14:58
  philosophy. If things are going well, keep
• 15:00
  going. If not, shake things up.
• 15:03
  It is, and that is essentially what these neural networks are
• 15:05
  doing is a process called gradient descent.
• 15:08
  They run a simulation. If the stress on the bracket is
• 15:11
  high, that's bad. So they twiddled the design
• 15:14
  They. Try something new.
• 15:15
  Right. If the stress goes down, that's
• 15:16
  good. So they keep going in that
• 15:18
  direction. They're just minimizing the
• 15:20
  error to evolve the perfect part.
• 15:22
  So that NASA bracket is basically just a very smart
• 15:25
  metallic soap bubble. In a mathematical sense, yes,
• 15:28
  it's finding the path of least resistance.
• 15:30
  So what does this all mean for our listeners?
• 15:32
  I mean, we've gone from NASA's twin spaceships keeping
• 15:35
  astronauts safe to AI that grows parts like bones, and now we're
• 15:40
  facing the need to stay one step ahead of the wanna laughs of the
• 15:43
  world. I think the big conclusion is
• 15:45
  that the future isn't about AI replacing engineers, it's about
• 15:49
  engineers evolving, levelling up.
• 15:52
  Exactly. The value of an engineer is
• 15:55
  shifting. It's moving from the act of
• 15:58
  drawing up the manual task of creating geometry to the act of
• 16:02
  understanding. So the engineer of the future is
• 16:05
  that architect of constraints. That's it.
• 16:08
  The better you can describe the problem, the loads, the heat,
• 16:12
  the materials, the better the AI can solve it.
• 16:14
  It's not about having the answers anymore, it's about
• 16:16
  asking the right questions. Spot on.
• 16:19
  If you could define the constraints, you control the
• 16:21
  output. That is a lot to think about.
• 16:23
  Whether you're designing A spacecraft or just trying to
• 16:25
  organize your wiring harness, the tools are changing the game.
• 16:28
  Absolutely. If you haven't already, get over
• 16:30
  to our YouTube channel Mechanical Engineering Made
• 16:33
  Simple, where we post hands on how to videos, visual
• 16:36
  breakdowns, and practical engineering walkthroughs.
• 16:38
  You can actually actually use a family that solves together
• 16:40
  stays together. Thanks for listening.