Latest / Mechanical Engineering Made Simple / (#24) Thermal Design Betrayals: How Engineers Battle Fouling, Hysteresis, and Phonon Mismatch to Build Reliable Systems

Transcript

(#24) Thermal Design Betrayals: How Engineers Battle Fouling, Hysteresis, and Phonon Mismatch to Build Reliable Systems · Mechanical Engineering Made Simple

• 0:00
  Welcome back, design junkies, to mechanical engineering made
• 0:02
  simple. If you've ever looked at a
• 0:04
  perfectly designed piece of equipment, say a massive
• 0:07
  industrial heat exchanger, and thought that looks simple, well,
• 0:10
  today's deep dive is here to maybe shatter that illusion a
• 0:14
  little. We are not just looking at the
• 0:17
  theory of how heat should flow, you know.
• 0:19
  We are fighting the Engineer's eternal battle against real
• 0:23
  world imperfection. That's absolutely right.
• 0:25
  Yeah. We're moving from the pristine
• 0:28
  world of theoretical thermodynamics to the messy
• 0:31
  reality where things breakdown, surfaces get gunked up,
• 0:34
  measurements lie. Our mission today is really to
• 0:37
  unpack the sources that detail how engineers quantify and
• 0:40
  control these factors. We're talking things like
• 0:42
  thermal decay, managing critical measurement errors, synthesizing
• 0:46
  mountains of imperfect experimental data into a
• 0:48
  reliable predictive model. Because in the field,
• 0:51
  performance isn't just about picking the right material.
• 0:53
  It's often dominated by factors like, well, geometry and just
• 0:57
  plain mechanical degradation over time.
• 0:59
  OK, so let's start with the ideal scenario first.
• 1:02
  We base pretty much everything in thermal analysis on the
• 1:05
  resistance analogy, treating heat flow like current in an
• 1:09
  electrical circuit. We define the general thermal
• 1:12
  resistance R as the temperature drop delta T divided by the heat
• 1:17
  flow Q. It's elegant, really.
• 1:19
  It is, and for simple one-dimensional conduction it's
• 1:23
  beautifully defined R = L divided by the product of K * A.
• 1:27
  That's length L over conductivity K times area A.
• 1:31
  That is the perfect foundation for say a simle slab of
• 1:34
  material, nice and clean. But almost immediately we start
• 1:37
  intentionally comlicating things right, Usually to boost
• 1:40
  erformance. The easiest way to increase heat
• 1:42
  flow is to dramatically increase that surface area a usually by
• 1:46
  adding fins. Fins are a wonderful tool,
• 1:48
  absolutely, but they introduce the first major, let's say
• 1:52
  betrayal of the ideal model. Heat has to travel along that
• 1:55
  fin, and as it does, the temperature gradient drops.
• 1:57
  This means the fin out at its tip isn't transferring heat
• 2:00
  nearly as efficiently as the base, and this forces us to
• 2:03
  introduce the concept of overall surface efficiency.
• 2:06
  ADA sub OF. So it's not enough just to
• 2:08
  maximize the surface area, we have to account for how
• 2:10
  effective that area actually is. Precisely, Yeah.
• 2:13
  Our sources show the outside surface efficiency ETA sub of
• 2:17
  comma O is calculated by looking at the fraction of the surface
• 2:20
  that is finned's sub fo relative to the total surface area's sub
• 2:25
  O, and then multiplying that relationship by the fin
• 2:28
  efficiency itself, ETA sub fo. The core insight here is that a
• 2:33
  fin must be optimized for size, material, and geometry.
• 2:37
  You know, a gigantic fin made of the wrong material or one that's
• 2:40
  just too long can actually be a net negative because it's
• 2:43
  efficiency just plummets. So you're just hauling around
• 2:45
  dead weight at that point? Exactly.
• 2:47
  Dead weight that isn't doing its job properly.
• 2:48
  OK, now let's talk about what feels like the true enemy of
• 2:51
  long term design fouling. This is where those beautiful
• 2:55
  lab results go to die, isn't it? Oh, absolutely.
• 2:59
  Fowling is catastrophic, really, because it's the inevitable real
• 3:02
  world accumulation of insulating substances.
• 3:05
  You know, scale grime sludge that constantly degrades the
• 3:09
  system's performance over its life.
• 3:11
  Engineers must design the component to survive years of
• 3:14
  operation, meaning they have to factor in a potentially massive
• 3:16
  performance penalty right from day one.
• 3:18
  How? Significant.
• 3:19
  Are we talking here? Is it a minor tweak or?
• 3:21
  Oh, it's enormous. Often we are essentially an
• 3:24
  insulating layer whose resistance is often far greater
• 3:28
  than the actual film resistance of the fluid itself.
• 3:31
  I mean, look at some of the data points from the sources process
• 3:34
  liquid streams like say a caustic solutions.
• 3:36
  They show fouling resistance is around 3.5 * 10 to the power of
• 3:40
  -4 square meters Kelvin per Watt.
• 3:43
  But if you're dealing with something like heavy fuel oil,
• 3:45
  that resistance can jump way up maybe to 12.3 * 10 to the power
• 3:49
  of -4 square meters Kelvin per Watt.
• 3:51
  That's that's a huge design challenge, almost an order of
• 3:55
  magnitude in some cases. Wow.
• 3:56
  And when you have these geometry changes like fins or this
• 4:00
  dynamic degradation like fouling, the heat transfer
• 4:03
  coefficient is definitely not uniform across the entire
• 4:06
  surface anymore. So how do we calculate the
• 4:09
  performance of the whole heat exchanger then?
• 4:11
  You can't just use the simple log mean temperature difference,
• 4:14
  can you? No, you really can't, not
• 4:16
  reliably. We have to employ tools that
• 4:19
  handle that non uniformity and this is where the effectiveness
• 4:22
  NTU method or epsilon NU 2 method comes in.
• 4:25
  Right, the epsilon, yeah. It's a necessary analytical tool
• 4:28
  that calculates performance based on how close the heat
• 4:31
  exchanger gets to its maximum theoretical performance, which
• 4:35
  is crucial when film coefficients or fouling are
• 4:38
  varying significantly because of the geometry or the process
• 4:41
  conditions changing. It gives us a reliable way to
• 4:44
  design these things, even when the conditions are complex and,
• 4:47
  frankly, messy. OK.
• 4:48
  So we've established that physical reality is imperfect.
• 4:50
  The margins we build in for things like fouling and fin
• 4:53
  losses are only as good as the measurement we take to begin
• 4:55
  with. But trusting the thermometer or
• 4:57
  whatever sensor we're using, that's its own engineering
• 5:00
  challenge, right? It absolutely is.
• 5:03
  Measurement at its core is a relative effort.
• 5:06
  It's all about comparison against some agreed upon
• 5:08
  standard. Our anchor point for
• 5:10
  temperature, for example, is the international temperature scale
• 5:13
  of 1990, or IT 90. It's defined by fixed points,
• 5:17
  like the triple equilibrium state of water at precisely
• 5:20
  27316 Kelvin. Everything else relates back to
• 5:24
  that fundamental point. But the very sensor we used to
• 5:27
  compare the temperature can betray us.
• 5:28
  I'm thinking specifically about hysteresis in thermocouples.
• 5:31
  That sounds like a nightmare. It is.
• 5:33
  It's the ultimate measurement artifact, really.
• 5:35
  Hysteresis means the sensor output depends not just on the
• 5:37
  temperature now, but on its immediate history where it's
• 5:40
  just been. So for a thermocouple, you might
• 5:42
  find one sensitivity curve, let's say 1.5 millivolts per
• 5:47
  degree Celsius when the temperature is rising.
• 5:49
  Pretty straightforward. But when the temperature begins
• 5:51
  to fall, the sensitivity changes, and often nonlinearly.
• 5:54
  It might follow a completely different complex function like
• 5:57
  St. equals 1.5 * ( 2.33 - 0.13 three t + 0.0033 ^2 ) millivolts
• 6:06
  per degree Celsius. Wait, hang on.
• 6:08
  So you're saying we have to calculate if the device is
• 6:10
  heating up or cooling down just to trust the number coming off
• 6:14
  the sensor? To be truly precise, yes.
• 6:17
  Wow, that sounds like overkill until you realize that
• 6:20
  calibration error could completely sink the whole
• 6:22
  process analysis. Exactly right.
• 6:24
  Precision demands that level of detail sometimes.
• 6:27
  So once we gather a collection of these complex, carefully
• 6:30
  calibrated measurements, we then have to quantify our confidence
• 6:34
  in them using statistics. We define the deviation D sub I
• 6:39
  of any single measurement T sub I from the average T bar.
• 6:43
  Pretty standard and we use the standard deviation Sigma to
• 6:46
  quantify the precision and the spread of all our measured
• 6:50
  values. OK, makes sense.
• 6:51
  But what happens when we get a measurement that just looks
• 6:53
  wrong? You know, a wild outlier that
• 6:55
  throws off the whole data set? We can't just toss it out
• 6:57
  because we feel like it can. We no, you definitely need
• 7:00
  mathematical justification for that.
• 7:02
  And that's where chauvinist's criterion comes into play.
• 7:04
  This is kind of the mathematical defense you give your boss when
• 7:07
  you throw out that one data point that would have ruined the
• 7:10
  entire project budget or conclusion.
• 7:12
  OK, the criterion is based on a probability rule.
• 7:16
  Essentially, the probability of obtaining a certain deviation by
• 7:20
  chance should not be smaller than 1 divided by tuck in, where
• 7:24
  N is your total number of measurements.
• 7:27
  Can you give us a concrete example of that?
• 7:28
  Like what's the barrier? Sure.
• 7:30
  Let's say you perform 10 measurements, so n = 10.
• 7:33
  The criterion dictates that the maximum acceptable deviation
• 7:36
  from the mean is 1.96 times the standard deviation Sigma.
• 7:41
  So if a data point falls outside that 1.96 Sigma boundary, you
• 7:45
  have the statistical authority based on this criterion to
• 7:48
  reject it as a questionable outlier.
• 7:49
  It protects the integrity of your experiment.
• 7:51
  Got it. So it's a justifiable way to
• 7:53
  clean the data. OK, so we have clean
• 7:55
  statistically validated data points, but often the final
• 7:58
  result we care about, let's call it Q, depends on inputs from
• 8:01
  multiple sensors, say T1T2 and T3.
• 8:05
  How do we track how the tiny uncertainties from each of those
• 8:08
  inputs combine and affect the final answer?
• 8:10
  Right, this is the problem of uncertainty propagation, and the
• 8:13
  Klein and Mcclintock method provides the rigorous framework
• 8:16
  for handling it conceptually. The idea is to make sure that
• 8:20
  your overall uncertainty in the result you sub Q is calculated
• 8:23
  by rigorously combining the individual measurement
• 8:26
  uncertainties. You say bye.
• 8:28
  You calculate this using a square bond of the summation of
• 8:31
  terms, where each term basically accounts for how sensitive the
• 8:34
  final result Q is to a small change in that specific
• 8:37
  measurement T sub I scaled by the uncertainty U sub I of that
• 8:41
  measurement itself, all squared. OK, so it sounds complex
• 8:44
  mathematically, but the goal is to avoid underestimating the
• 8:47
  total error. Exactly.
• 8:48
  It prevents engineers from being overly confident and drastically
• 8:51
  underestimating the error bars on their final calculated
• 8:54
  answer. Super important.
• 8:56
  That makes sense. So once we trust the combined
• 8:59
  uncertainty, we usually need to convert that reliable data set
• 9:02
  into something usable, like a predictive tool, an equation, or
• 9:05
  a curve fit and. Precisely, if you measure, say,
• 9:09
  a fluid temperature T versus the heating power Q dot, you
• 9:12
  typically use the least squares method to find the line or curve
• 9:16
  that best represents that relationship.
• 9:18
  It works by finding the coefficients for your proposed
• 9:20
  function, maybe a straight line, maybe a polynomial, by
• 9:24
  minimizing D which is the cumulative sum of the squares of
• 9:28
  the distances between your measured points and the fitted
• 9:31
  line, right? So it doesn't let one wacky
• 9:33
  point skew the whole line too much.
• 9:35
  Exactly. And crucially, when calculating
• 9:37
  the standard error of that curve fit epsilon, you have to use n
• 9:41
  -, P in the denominator. N is the number of data points
• 9:44
  and P is the number of coefficients you determine for
• 9:46
  your fit. This properly accounts for the
• 9:48
  degrees of freedom you used up in finding that best fit line.
• 9:52
  And this whole rigorous process, careful measurements,
• 9:55
  statistical validation, uncertainty propagation, and
• 9:58
  then curve fitting, this is how we get those predictive rules
• 10:01
  that really underpin practical heat transfer design.
• 10:04
  Right. Yes, absolutely.
• 10:06
  This is how experimental measurements are synthesized
• 10:08
  into generalized, often dimensionless convective
• 10:11
  correlations. We eventually arrive at
• 10:13
  equations like the one for the Nusselt number nu, which equals
• 10:17
  some constant C times the Reynolds number RE raised to the
• 10:21
  power of N times the Prantle number PR raised to the power of
• 10:24
  M. This relationship, derived from
• 10:27
  countless experiments fitted this way, tells us how heat
• 10:30
  transfer scales with fluid speed.
• 10:32
  That's the RE part, and fluid properties the PR part in a
• 10:35
  predictable, experimentally verified way.
• 10:38
  It's powerful. But how do we validate the model
• 10:40
  itself? We verified the individual data
• 10:43
  points using chauvinist maybe, but how do we prove that the
• 10:46
  distribution of our experimental data actually aligns with the
• 10:49
  theoretical model or curve we created?
• 10:51
  Is the fit actually meaningful? Good point.
• 10:54
  That's where the chi square test comes in.
• 10:56
  It's a really foundational statistical test.
• 10:59
  It essentially compares the expected distribution of data
• 11:02
  points based on your theoretical model or curve fit against the
• 11:06
  observed distribution, what you actually got in your real world
• 11:09
  experiment. It helps engineers confirm
• 11:12
  whether the model they built is truly representative of the
• 11:14
  underlying physical phenomena, or if maybe they're just drawing
• 11:18
  a line through random noise that doesn't really follow the
• 11:20
  pattern they think it does. It validates the goodness of
• 11:23
  fit. OK, that makes sense.
• 11:24
  It's like a final check on the whole modeling process.
• 11:28
  Let's apply some of this precision thinking now to what
• 11:30
  feels like the modern battlefield of thermal
• 11:33
  engineering, electronic cooling. When we're trying to cool tiny
• 11:36
  powerful chips, we face resistance at interfaces.
• 11:39
  And you mentioned there are sort of two main types, macroscopic
• 11:42
  and microscopic. Correct.
• 11:43
  The macroscopic resistance is the one most people think of
• 11:45
  first. Thermal contact resistance are
• 11:48
  subco. This is caused by surface
• 11:51
  disparities, basically the microscopic roughness and
• 11:54
  imperfect physical contact. When you press 2 surfaces
• 11:56
  together, even seemingly smooth surfaces only touch at a few
• 12:00
  points. Right.
• 12:01
  And that's what we use thermal interface materials for, to fill
• 12:04
  those gaps. Exactly.
• 12:05
  Then there's the microscopic resistance, which is maybe less
• 12:08
  intuitive, the thermal boundary resistance R sub B.
• 12:12
  This occurs even at a nearly perfect atomically flat
• 12:15
  interface. OK R sub B.
• 12:18
  Why does that exist? If the surfaces are almost
• 12:20
  touching perfectly, what's resisting the heat flow them?
• 12:24
  It's down to physics at the atomic level, really.
• 12:26
  Heat and solids is primarily carried by packets of
• 12:29
  vibrational energy called phonons.
• 12:31
  Think of them like sound waves, but at much higher frequencies.
• 12:34
  R sub B exists because the phonons in chip material A
• 12:37
  literally don't speak the same vibrational language as the
• 12:40
  phonons in material B. The heat sing material, for
• 12:43
  instance. There's an acoustic mismatch.
• 12:45
  Those thumbel vibrations don't efficiently pass across the
• 12:47
  boundary because the materials have different vibrational
• 12:50
  properties. This creates resistance.
• 12:52
  It's usually only significant at the actual transistor level, or
• 12:55
  maybe at very low temperatures, but it illustrates the
• 12:58
  incredible depth of the complexity we deal with.
• 13:01
  Fascinating. OK, so back to the macroscopic R
• 13:04
  subco. To combat that we use thermal
• 13:07
  interface materials or Tim's off of these are particle even
• 13:10
  polymers designed specifically to fill the air gaps.
• 13:13
  But I remember reading that using the Tim material with the
• 13:16
  absolute highest bulk thermal conductivity can actually
• 13:19
  backfire. Is that right?
• 13:20
  That's one of the greatest counter intuitive trade-offs in
• 13:22
  thermal design. It really is.
• 13:24
  Tim's are often filled with highly conductive particles like
• 13:27
  boron nitride or diamond dust to increase their bulk thermal
• 13:31
  conductivity. Kappa sub Tim makes sense,
• 13:34
  right? Yes, seems obvious.
• 13:35
  More conductive filler, Better heat transfer.
• 13:37
  But adding more of these hard particles also increases the
• 13:40
  mechanical rigidity or the shear modulus of the polymer matrix
• 13:44
  material itself. And increased rigidity means the
• 13:47
  material resistance being deformed, so it leads to a
• 13:50
  thicker bond line delta when you assemble the parts.
• 13:53
  It also means a smaller reel contact area, a sub reel,
• 13:56
  because the stiff material can't conform as well to the
• 13:59
  microscopic surface roughness. So the material becomes too
• 14:02
  stiff to properly squish into all those tiny surface crevices
• 14:06
  and make good contact. Exactly.
• 14:08
  It resists being squeezed into that ultra thin layer you
• 14:11
  ideally want because of this conflict.
• 14:14
  Better bulk conductivity versus worse mechanical conformance.
• 14:17
  The minimum overall thermal resistance, which means the best
• 14:20
  thermal performance might actually occur at particle
• 14:23
  loadings that are less than the maximum possible concentration.
• 14:26
  Wow, yeah you can literally over engineer the materials bulk
• 14:29
  properties and end up the critical mechanical contact
• 14:33
  needed for the thermal path. It's a balancing act.
• 14:35
  OK, that is counterintuitive. Speaking of high performance, we
• 14:39
  have to touch briefly on heat pipes.
• 14:41
  These things seem almost like magic in their efficiency
• 14:44
  sometimes. They are phenomenal devices.
• 14:46
  Yeah, they operate using phase change liquid evaporating in one
• 14:49
  section, vapor travelling and then condensing in another
• 14:52
  section, all driven by capillary forces within a Wick structure.
• 14:56
  Their effective axial thermal conductivity is truly
• 14:59
  astonishing. A simple heat pipe using water
• 15:02
  can reach effective conductivities around, say
• 15:05
  100,000 watts per meter per Kelvin along its line A.
• 15:08
  100,000. Yeah.
• 15:09
  Compare that to solid copper, which is one of our best
• 15:12
  conductors at only around 400 watts per meter per Kelvin.
• 15:15
  It's orders of magnitude difference.
• 15:17
  Just incredible devices. And finally, let's look at a
• 15:20
  really cutting edge technique, immersion cooling using these
• 15:24
  special low boiling point dielectric fluids.
• 15:27
  Highly efficient potentially, but it introduces this massive
• 15:30
  stability issue called incipients overshoot.
• 15:32
  What's going on there? Right incipients overshoot is a
• 15:35
  serious headache. It's basically an unwanted
• 15:38
  thermal shock caused by delayed boiling.
• 15:41
  These dielectric fluids, like FC-72 for example, have a
• 15:44
  surprisingly high solubility for air.
• 15:47
  They literally dissolve the air out of the tiny microscopic
• 15:50
  cavities that naturally exist on the chip surface.
• 15:53
  OK. Why does that matter?
• 15:54
  Because those air filled cavities are normally the
• 15:57
  nucleation sites, the easy starting points for boiling to
• 16:00
  begin when they're gone, filled with liquid instead of air, it
• 16:04
  takes a much higher surface temperature, a higher superheat,
• 16:07
  to initiate boiling. So the surface gets hotter and
• 16:10
  hotter than boiling finally does initiate, but it happens almost
• 16:14
  violently and suddenly over a large area.
• 16:16
  This causes an undesirable spike in temperature, followed by a
• 16:19
  sudden jarring drop as the vigorous boiling quickly cools
• 16:23
  the surface. So that initial spike followed
• 16:25
  by the rapid crash, it is the overshoot.
• 16:28
  That's the overshoot, and that thermal cycling that shock is a
• 16:31
  major design challenge for long term component reliability.
• 16:34
  You don't want your chip going through those rapid temperature
• 16:37
  swings repeatedly. Definitely not.
• 16:39
  OK. So to sort of tie all these
• 16:42
  complexities together, maybe we end where engineers often start
• 16:47
  their transient analysis with a critical simplification tool,
• 16:52
  the BIOT number. Ah yes, the BIOT number Bo.
• 16:55
  It's fundamental because it essentially gives you permission
• 16:58
  to simplify reality. Sometimes.
• 17:00
  It's the ratio of the internal conduction resistance within a
• 17:03
  solid body to the external convection resistance at its
• 17:06
  surface. Mathematically, Bi equals H * L
• 17:10
  / K, where H is the convection coefficient, L is a
• 17:13
  characteristic length, and K is the thermal conductivity of the
• 17:16
  solid. And the magic number is 0.1.
• 17:19
  Pretty much, if the BIOT number is less than about .1, it means
• 17:22
  the internal conduction resistance is negligible
• 17:24
  compared to the resistance to heat getting away from the
• 17:26
  surface. This implies that the
• 17:28
  temperature inside the body is essentially uniform at any given
• 17:31
  time, and if that's true, we can use the much simpler lumped
• 17:34
  capacity solution for transient analysis, treating the whole
• 17:37
  object as being at a single temperature that changes
• 17:39
  overtime. So it's the metric that saves
• 17:41
  you from a potential lifetime of solving unnecessarily complex
• 17:46
  partial differential equations for internal temperature
• 17:49
  gradients. You got it.
• 17:50
  It tells you when you can ignore those internal details and still
• 17:54
  get an accurate answer for the overall cooling or heating
• 17:57
  behavior. It's a hugely important
• 17:59
  simplification tool. So we've really traveled quite a
• 18:02
  path today from the ideal foundation of thermal resistance
• 18:07
  through the sort of statistical rigor needed to trust even a
• 18:10
  single data point, into the messy material driven world of
• 18:13
  fouling and this sometimes counterintuitive design of
• 18:16
  things like Tim's. We've seen that modern
• 18:18
  engineering isn't just about applying a perfect equation from
• 18:20
  a textbook, it's really about wielding this whole toolbox.
• 18:23
  Things like the BIOT number, the epsilon 2 method for exchangers,
• 18:27
  statistical validation like the chi square test to manage, model
• 18:30
  and ultimately survive real world imperfection.
• 18:33
  And what's really fascinating when you step back is realizing
• 18:35
  that pretty much every engineered component is a
• 18:38
  complex collection of compromises, isn't it?
• 18:41
  I mean, just think about something like a printed circuit
• 18:43
  board, considering the incredibly complex interaction
• 18:46
  between the thermal V is trying to pull heat out the anisotropic
• 18:49
  conductivity of the board materials themselves.
• 18:51
  All those different trace layers for electricity.
• 18:54
  How many layers upon layers of interconnected thermal and
• 18:57
  electrical compromises are embedded just in the structure
• 19:00
  of the devices you use every single day?
• 19:03
  And maybe more importantly, how many distinct analytical checks,
• 19:06
  experimental validations, and statistical tests like the ones
• 19:09
  we discussed were required just to ensure those devices function
• 19:13
  reliably without melting? It's quite staggering when you
• 19:16
  think about it. Until the next equation design
• 19:18
  junkies, we are mechanical engineering made simple.