Hand-considered hypotheses, 4301–4400
Semiconductor packaging and device reliability, written one at a time. The attach under the die, the wire and its three joints, the moulding compound as a mechanical and a chemical participant, the thermal path and the instruments that measure it, the solder joint at board level, the discharge event, moisture inside a sealed volume, and what failure analysis does to the sample it examines. Each entry carries the logical form chosen for it, printed above the claim, so the variety of argument can be audited against the text rather than taken on trust. None has been tested; no citation is given because none is being invoked, and every quantity offered as an estimate is marked as one.
Die attach and the die-level interface
Misattribution
H4301. An attach void that appears to grow during thermal cycling in acoustic imaging is often the die's backside metallization lifting from the silicon, and the two cannot be told apart acoustically because both present as a reflecting interface of the same polarity. A die-shear test separates them, since backside separation leaves the shear strength almost unchanged while a true attach void reduces it in proportion to the lost area. The estimate worth testing is that a large share of what is called attach voiding in field returns is this backside separation.
Hidden term
H4302. Every die-attach thermal resistance quoted from a structure function contains the die's own spreading resistance under the pad, a term set by the die thickness and the heat-spreader footprint rather than by the attach. Because that term is fixed for a given die, comparing two pastes on dies of different thickness ranks them by the wrong quantity. Thinning one die in steps and re-measuring isolates the term without changing the attach.
Compliance claim
H4303. A sintered silver layer with higher porosity can outlast a denser one at the same thermal resistance, because the voids give the layer somewhere to strain and the denser layer puts that strain into the silicon. Matching the two by resistance requires changing thickness, so the test is a pair of pastes whose resistances coincide while their porosities do not. If the claim holds, the industry's direction of travel toward denser sinter is backwards for large dies.
Bound
H4304. There is a minimum bond-line thickness for any filled die attach, set by the largest filler particle rather than by the press force, and below it the resistance rises no matter how hard the die is pressed. The bound is geometric, because particles bridge the gap and hold the die off the leadframe, so extra force only loads the die; grinding the paste finer is what lowers the minimum at a fixed pressure.
Sign of the trend
H4305. A soft die attach ordinarily ages toward higher thermal resistance, but one that creeps under thermal cycling can thin and give a resistance that falls while its stress rises. The fall is a warning rather than an improvement, and it should appear in the structure function as a shift of the attach peak toward shorter times. Any qualification that tracks only resistance will pass the part on its way to the worst state.
Identity of two modes
H4306. Delamination at the die-attach interface and cohesive cracking through the attach are reported as two failure modes, but in a sintered layer both follow the same density minimum, which sits within a short distance of the interface, by estimate a tenth of the layer thickness. A density map across the layer and a crack path on the same specimen should coincide if that is so. If they do, the two modes are one gradient and the process fix is the same for both.
Threshold
H4307. There is a die size above which a given attach thickness cannot absorb the silicon-to-leadframe expansion without the die seeing more tensile stress than it can take. A useful estimate is that the threshold grows with the square root of the thickness, so doubling the attach buys roughly one and a half times the die edge. A size series at fixed thickness should find the threshold as a sharp onset rather than a gradual trend.
Inversion of standard practice
H4308. The standard instruction to make the bond line as thin as possible is right for thermal resistance and wrong for reliability, because below a thickness the joint's compliance is gone and the silicon carries the mismatch. The optimum is where the resistance gained per micrometre removed equals the stress it costs, and that point is thicker than most specifications allow. Cycle a thickness series to first crack and the thin end is where life collapses, not where it merely declines.
Definitional error
H4309. The stress-free temperature assigned to a die attach is not a material property but the temperature at which the attach's relaxation time matches the cool-down time, so it depends on the cooling rate used to measure it. Slow cooling gives a low apparent stress-free temperature and a low measured die stress, while fast cooling gives the opposite. Two laboratories using the same paste and different profiles will disagree, and neither will be wrong.
Order-of-limits swap
H4310. Selecting an attach by thermal resistance first and checking mechanical stress second gives a different answer than doing the two in the other order, because the mechanical check changes the geometry for which the thermal number was computed. The correct procedure is a fixed point rather than a sequence, and a single pass can choose a paste that fails the check it never re-ran. Any selection list built in one direction should be rebuilt in the other.
Scaling collapse
H4311. Warpage from dies of different thickness and attaches of different thickness should collapse onto one curve when plotted against the ratio of the two, because that ratio alone sets how much of the mismatch the attach absorbs. If the points do not collapse, the attach is not the compliance in the stack and the search should move to the leadframe; the test costs three die thicknesses, three attach thicknesses, and one plot.
Two-mechanism separability
H4312. Moisture preconditioning and thermal cycling both raise an attach's thermal resistance by a similar amount, but they do it in different places, since moisture attacks the interface as a thin resistive layer while cycling damages the bulk. The transient structure function separates them by the position of the new peak rather than by its height. A qualification that reports only the steady resistance cannot say which damage it has induced.
Non-monotonic window
H4313. Adhesion promoter at the die-attach interface improves wet adhesion up to roughly monolayer coverage and then plasticises the interface beyond it, so shear strength passes through a maximum instead of reaching a plateau. The window is narrow enough that a formulator working by weight fraction can cross it without noticing. A concentration series at constant cure places the maximum, and its position is the useful number.
Latency claim
H4314. An attach that looks intact in acoustic imaging can hold a crack that is closed at room temperature and opens only when the die is powered and the joint is hot. The population of no-fault-found returns is therefore enriched in parts whose only defect is invisible until the operating condition is applied. The test is powered thermal imaging after cycling, not before.
Saturation claim
H4315. Die shear strength saturates with cure time before the glass transition does, so a schedule validated by shear stops short of full cure and leaves a residual that a moisture test later finds. The two properties have different time constants because one is set by crosslink density at the interface and the other by the bulk network. Extending the cure until the transition stops moving, then re-checking shear, exposes the gap.
Load-path claim
H4316. In a stacked-die package the attach under the top die carries almost none of the thermal load and most of the mechanical load, so optimising every layer for conductance misdirects the effort. The allocation that follows is thermal paste at the bottom die with the compliant film at the top, and measuring the layers separately in a two-die stack shows the asymmetry at once.
Boundary-condition error
H4317. Models of die stress in an attached die usually leave the top and edges traction-free, but the moulding compound bonds there and adds a constraint that can cut the peak tensile stress by an estimated third. The error is conservative for void growth and unconservative for die cracking, because the constraint also changes the stress triaxiality. A model run with the compound removed should disagree with a measured die curvature by about the size of that term.
Ratio invariant
H4318. The ratio of die shear strength after moisture soaking to that before is nearly the same for all epoxy attaches, because the loss tracks the water concentration at the interface and every epoxy reaches a similar concentration at saturation. Ranking pastes by wet strength is therefore ranking them by dry strength with a common multiplier applied. Soak three chemically different pastes and the ratios cluster while the absolutes do not.
Constraint argument
H4319. Beyond a certain die size no single attach can be both the thermal and the mechanical solution, because the property that lowers resistance, stiffness, is the one that transmits stress. The constraint is geometric rather than a materials gap, and it is why large dies migrate to a two-material stack. The crossover should appear as a die size at which no paste in a family passes both screens.
Necessity claim
H4320. A void-free attach is necessary for thermal resistance and not for reliability, and conflating the two has produced void specifications with no physical basis. The reliability-relevant quantity is the largest void's size relative to the heat-spreading length under the hotspot, not the total void fraction. A specification written as a fraction should be replaced by one written as a length, and the two rank pastes differently.
Wire bonds and the interconnect
Rival cause
H4321. Neck breaks after thermal cycling are usually charged to intermetallic growth, but the neck is also where ultrasonic bonding has worked and recrystallised the wire, and that zone is weaker than the intermetallic. Two wires with the same intermetallic thickness and different bonding energy should fail differently, which separates the explanations. My estimate is that most of the ductility lost at the neck is mechanical history rather than diffusion.
Stiffness contribution
H4322. A wire pull test measures the wire as well as the bond, and for a short loop the wire's own bending stiffness is a large fraction of the reading, so pull values from short and long loops are not comparable and the difference is a property of neither bond. Bonding adjacent pads at two loop heights and pulling both is the practical way to isolate the term.
Test-created artefact
H4323. Bond lift in a pull test can be manufactured by the test itself, because if the wire is stiffer than the bond the hook lifts the ball before the wire breaks and the failure is recorded as an interfacial weakness. A shear test on the same bond either finds a normal force or finds nothing, which distinguishes a weak bond from a strong one tested the wrong way. Every bond-lift rate should be checked against shear before it is believed.
Optimum on the damage side
H4324. Raising ultrasonic power strengthens a ball bond up to the point where the pad's dielectric stack begins to crater, and on thin low-k stacks that point can arrive before the bond is fully formed. The optimum therefore sits against the damage boundary rather than in the middle, which inverts the usual instruction to bond at the lowest power that passes. A power series read with both pull and crater inspection locates the overlap.
Pressure threshold
H4325. There is a minimum pressure below which an aluminium pad's native oxide is not broken and the bond forms only on asperities, so pull strength is set by contact area rather than by any metallurgical joint. The threshold is a pressure and not a force, which is why small pads bond better at the same force. Run a pad-size series at constant force and the transition is where contact stops being asperity-limited.
Mechanism identity
H4326. Copper-aluminium and gold-aluminium bond failures are described as different metallurgies, but both are the same volume mismatch, since the intermetallic occupies less volume than the metal it consumed and the crack opens on the wire side in each. The difference in rate is a difference in diffusion and not in mechanism. Cut both systems at matched intermetallic thickness and the crack sits in the same place.
Wake-shielding claim
H4327. Adding a second wire close beside a first reduces the first wire's sweep during moulding, because the pair shares a wake and the downstream wire meets a slower front. Denser arrays can therefore be more reliable in sweep than sparse ones, which is the opposite of the usual density concern. Sweep the pitch at constant fill speed and the minimum appears at a spacing of a few wire diameters, by estimate.
Measurement drift
H4328. The loop height measured optically before moulding is not the loop height in service, because the wire sags between bonding and encapsulation and the sag depends on the span and on the annealing the wire received during bonding. Two wires bonded identically at different spans therefore do not sit at the heights the machine recorded.
Count argument
H4329. A wire bond has three independent failure sites, the ball-to-pad interface, the heat-affected neck, and the stitch heel, and a qualification that reports one number is averaging three different processes. The number of independent process knobs is also three, one per site, so an experiment with fewer factors cannot resolve them. Reporting the three separately costs nothing and makes the average meaningful.
Selection effect
H4330. Failure analysis selects the wire that broke, which is the one with the thinnest neck in the population, and generalising its strength to the lot is a selection effect in its purest form. The surviving wires are systematically stronger because they had to survive the same test. Pulling a random sample from the same lot, rather than the failed unit, is the only way to get the population value.
Criterion claim
H4331. The quality criterion for a stitch bond should be the stitch's departure angle rather than its pull force, because the force is dominated by the wire's diameter while the angle comes from the tool's trajectory and cannot be raised by using a stronger wire. A bond with a shallow angle fails at the heel later, and the force gives no warning, so an angle specification would be harder to meet and more informative.
Fracture scaling
H4332. The pad-cratering threshold scales with the dielectric stack's thickness and fracture toughness and not with the pad's area, so smaller pads are not automatically safer and larger pads are not automatically worse. What changes with area is the force needed to reach the threshold pressure. Hold the pressure constant across a pad-size series and the crater threshold stops moving.
Signature pair
H4333. The darkening seen at gold-aluminium bonds has two causes, intermetallic formation and void coalescence, and they differ in their spatial pattern rather than in their colour. Intermetallic growth is uniform over the bonded area because it follows temperature, while voiding concentrates where current crowds and therefore appears patchy. Stress one bond with current and leave its twin at the same temperature, and the patterns diverge.
Diameter optimum
H4334. Wire diameter has an optimum rather than a monotonic benefit, because a thicker wire carries more current but needs more force, and the force that deforms it craters the pad before the wire is fully bonded. The optimum is fixed by the pad stack and not by the current requirement. A diameter series at constant current density puts the life maximum in the middle.
Latent crack claim
H4335. A wire with a partly cracked heel can pass a pull test and then fail in service, because the moulding compound supports the loop and hides the crack until the compound itself moves. The moulding is load-bearing in that sense and not merely protective, and a qualification that tests bare wires tests a different structure. Pull moulded and unmoulded parts from one lot and the difference is the compound's contribution.
Rate separation
H4336. Bond strength saturates with bonding time long before cratering begins, so over-bonding is a damage risk and not a strength benefit, and the two have different time constants. Strength saturates when the contact area is complete, while damage accumulates as long as ultrasonic energy continues into the pad. A time series with strength and crater inspection together would locate the gap between them.
Sequence dependence
H4337. The order in which the two ends of a wire are bonded changes the residual stress at the heel even with identical machine settings, because the first bond fixes the wire's length and the second pulls against it. Ball-first and stitch-first are therefore different processes wearing the same recipe. Bond the same wire in both orders and read heel stress from curvature.
Closure argument
H4338. A wire bond's reliability is the minimum over its three joints, and a qualification that improves the average can lower the minimum by moving stress to the weakest one. The reportable number is the weakest joint's margin and not the mean pull. This is why a wire-bond improvement sometimes appears as a stitch failure with no change to the stitch process.
Moulding compound and the package body
Wrong heat source
H4339. Package warpage is usually charged to the moulding compound's expansion mismatch with the leadframe, but at thin body thicknesses a larger term is the cure shrinkage gradient through the thickness, which persists after the part is cool. The two are separable because shrinkage does not reverse on reheating while expansion does. A warpage pair taken above and below the cure temperature gives the split directly.
Third-material claim
H4340. Every adhesion value for a moulding compound on copper includes a layer of oxide that grew during the moulding's own heat, so the interface being measured is a three-material stack and not the one named. The oxide's thickness depends on the leadframe's storage and preheat, which is why adhesion data from two sites disagree. Measuring the same compound on clean and oxidised copper separates the interface from the oxide.
Wetting threshold
H4341. There is a filler content above which a compound stops wetting the leadframe, because the resin that must reach the surface is excluded by the packing of the filler network. The threshold is governed by the filler's surface treatment rather than by loading alone, so two compounds at the same loading can sit on opposite sides. A wetting-balance series at constant loading and varying treatment should find it.
Competing effects
H4342. Raising the silica loading lowers the compound's expansion and therefore looks protective, but it also raises the modulus, and past a loading the stress on the die rises even as the expansion match improves. The loading that matches expansion is not the loading that minimises die stress, and the second is lower. Two compounds of equal expansion and unequal modulus diverge in die stress.
Coverage optimum
H4343. Adhesion promoter improves the compound-to-leadframe bond up to monolayer coverage and then forms a weak boundary layer of unbonded silane, so adhesion passes through a maximum and falls. The fall is a reversal and not a saturation, and it is why over-dosing a promoter in production can look like contamination. A concentration series at fixed cure traces both the maximum and the decline.
Geometric artefact
H4344. The delamination that acoustic imaging reports at a leadframe tip is often the reflection from the leadframe's own step, because a step and a thin air layer have similar acoustic contrast. The artefact appears at the same place in every part of a lot, which is the tell. Imaging an unexposed control from the same lot distinguishes the geometry from the defect.
Naming duplication
H4345. The knit line and the flow line in a moulded package are the same feature seen at different fill stages, and the weakness attributed to a weld line is the filler-depleted band that forms wherever two fronts meet. Because it is one feature, the two names have supported two separate sets of process fixes for one problem. A short-shot series catches the band forming before either name is needed.
Instrument lag
H4346. The mould's temperature is controlled and measured at the platen, but the cavity lags by a time comparable to the cure, so the degree of cure at ejection is set by a temperature nobody records. Cycle-time changes attributed to cooling are therefore partly cure changes. Instrument the cavity and the lag appears, with its effect on eject strength.
Press bound
H4347. There is a package thickness above which the mould's clamp force cannot keep the two halves closed, so the limit on thin packages becomes the press rather than the compound. The bound scales with projected area and cavity pressure, so it arrives suddenly as body size grows and should appear as a step in flash rather than as a trend on a thickness series at one press.
Stiffness crossover
H4348. Warpage of a moulded package is dominated by the leadframe at small body size and by the compound at large size, and the crossover is where the two stiffnesses are equal. Below it, changing the compound does almost nothing, while above it, changing the leadframe does almost nothing. Two compounds and two leadframes across a body-size series cross at one point.
Process window
H4349. Vacuum-assisted moulding reduces voids up to a point and then pulls resin away from the surface, leaving a filler-rich skin and a resin-depleted layer that bonds poorly. There is therefore a vacuum window whose upper edge is a surface chemistry effect and not a flow effect. Measure void count and surface resin content together across a vacuum series and both edges appear.
Deferred cure
H4350. A moulding compound continues to cure during board assembly, so the properties measured after post-mould cure are not the properties the part has when it is soldered. The second cure is a process step that is being skipped, and its absence shows up as a shift in the compound's expansion at reflow rather than as a loss of strength. Measure one part before and after a simulated assembly excursion and the shift is the skipped cure.
Geometric saturation
H4351. Warpage saturates as a package is made thicker because the neutral axis moves out of the compound and into the leadframe, and the saturation thickness is a geometry rather than a material property. One compound across a thickness series gives the plateau and where it begins.
Modulus paradox
H4352. A softer moulding compound can protect the die better than a matched one, because the die's stress is governed by the compound's in-plane constraint and not by the expansion difference alone. Matching expansion with a stiff compound can therefore raise the stress it was meant to lower. Two compounds with equal expansion and different modulus should rank in the opposite order to the usual expectation.
Wrong state
H4353. Warpage measured on a free part is not the warpage that matters, because the socket constrains the part and that constraint sets the stress in the joints. A specification written on the free part can pass a package that fails when it is clamped. Measuring the same part free and in its socket would give the correction, and the difference should scale with the socket's contact stiffness.
Parallel shift
H4354. The ratio of a package's warpage at two temperatures is nearly the same for all moulding compounds, because in the glassy range the warpage follows the leadframe's expansion. Compound choice therefore sets the offset and not the slope, which is why warpage data from different compounds look parallel. Two temperatures and three compounds give the common ratio.
Proxy failure
H4355. The correct criterion for choosing a moulding compound is the product of its modulus and its expansion mismatch with the die, not the mismatch alone, and the two criteria can rank compounds in opposite orders. Expansion matching has become a proxy for stress, and the proxy fails for low-modulus compounds. Re-ranking a supplier's catalogue by the product would change the selection.
Weakest interface
H4356. A moulded package has three interfaces that can delaminate, die to compound, leadframe to compound, and die to attach, and its reliability is the weakest of them. Improving one moves the load to another, so a programme that fixes the reported failure can produce a different one at the same rate. The only durable improvement raises all three or moves the weakest far enough that another becomes the limit.
Thermal paths and their measurement
Board dominance
H4357. The junction-to-air resistance quoted for a package in still air is mostly a property of the board it is soldered to, because the board is the spreading surface and the package is a small perturbation on it. Two packages with the same number on different boards can differ by more than either differs from its neighbour. Measuring the same package on two board sizes separates the package from the board.
Self-heating term
H4358. A junction temperature inferred from a forward-voltage calibration includes the self-heating of the calibration current itself, and for a small die that term is not negligible. The calibration is measured at currents low enough to seem harmless and not low enough to be isothermal. Repeating it at two sense currents and extrapolating to zero removes the term.
Constriction claim
H4359. The heat leaving a die through its backside metal is constricted where it spreads into that metal, so the effective spreading depends on the metal's thickness relative to the heat source's width and not on the die's area. The constriction persists even when the metal is far more conductive than the silicon, because it is a geometric effect. A die with two backside thicknesses gives a resistance that does not scale with area.
Thickness paradox
H4360. Improving a thermal interface material's conductivity can raise the junction temperature, because a stiffer or more heavily filled grease resists being squeezed and leaves a thicker bond line whose penalty outweighs the conductivity gain. The effect is largest where the clamping pressure is low. Two greases of different conductivity at two pressures cross at the inversion.
Load-line threshold
H4361. There is a junction temperature above which a power device's operating point becomes unstable, and the threshold is fixed by the product of the thermal resistance and the temperature coefficient of the on-state loss. The instability belongs to the load line and not to the material, so the same die can be stable in one package and not in another. A power sweep at two case temperatures should find the threshold as a fold in the voltage-current locus.
Spectral identity
H4362. The peaks of a transient thermal structure function are usually read as physical layers, one peak per layer, but a single thick layer with distributed resistance and capacitance produces a broad peak that can be mistaken for two. The correspondence is not one to one, and the number of peaks is a property of the time resolution as much as of the stack. A known single layer measured at two resolutions splits into two apparent peaks.
Window artefact
H4363. The thermal resistance extracted from a transient measurement depends on the time window chosen, because a short window truncates the slow paths and returns a lower number. Two laboratories can therefore report different resistances for identical parts without either being wrong. Reporting the window with the number, or extending it until the curve's slope is constant, would make the values comparable.
Inflection claim
H4364. At low power the junction temperature rises sublinearly with power because spreading improves, and at high power it rises superlinearly because the silicon's conductivity falls, so the curve has an inflection that sets the usable range. The inflection is not a failure threshold but the point where two trends exchange dominance. A power sweep with the case held at two temperatures should locate it.
Spreading bound
H4365. There is a minimum thermal resistance per unit footprint set by spreading in the heat spreader, and below that bound added copper or added vias do not help because the heat has nowhere to go. The bound comes from the spreader's thickness and conductivity rather than from the die, so a spreader-thickness series at fixed footprint should approach it and stop.
Density scaling
H4366. An array of thermal vias conducts in proportion to the number of vias once the pitch is shorter than the spreading length in the surrounding material, so diameter matters much less than count. Below that pitch the array behaves as a homogeneous layer and the design variable is density. Two arrays with equal copper fraction and unequal pitch differ in conductance.
Thickness optimum
H4367. Die thickness has an optimum for junction temperature, because a thicker die spreads heat over a larger area while adding path length, and the two effects cross at a thickness comparable to the heat source's width. Thinning below the optimum helps only when the source is smaller than the die. Fix the power, vary the thickness, and the optimum moves with the source size.
Mismatched time constants
H4368. The package's thermal time constant is long compared with a short electrical overload, so the junction does not respond while the solder joint does, and the joint becomes the recorder of an event the thermometry missed. This is how an overload leaves a cracked joint and a normal junction temperature. A short pulse that leaves the junction cool and cracks the joint is the demonstration.
Hotspot criterion
H4369. The right thermal criterion for a die attach is the peak temperature at the hotspot rather than the average resistance across the die, and the two rank materials differently when the hotspot is small compared with the die. A material with high average conductance and poor local spreading can lose to one with the opposite. A hotspot of controlled size, moved across the die, changes the ranking.
Cancellation trap
H4370. The fall of carrier mobility with temperature and the rise of a package's thermal resistance with temperature can cancel over a range, making a device look thermally stable exactly where it is accumulating margin. The cancellation temperature belongs to the pair and not to the device, and it moves when the package changes. Comparing two packages of different resistance locates the cancellation without any device model.
Stack confound
H4371. A measured junction-to-case resistance includes the interface material between the package and the cold plate, and no single measurement can separate the two. The number belongs to the test stack and not to the package, which is why values differ between laboratories with different fixtures. Two fixture pressures would give the slope of the interface term and allow the intercept.
Moving bottleneck
H4372. Every thermal path has a bottleneck, and the bottleneck moves with power because spreading resistance depends on conductivity and conductivity depends on temperature. A design optimised at one power is therefore not optimised at another, and the optimum shifts toward more spreading as power rises. Two spreader designs across a power series cross where the bottleneck changes.
Solder joints and board-level assembly
Upstream cause
H4373. Solder-joint fatigue in board-level tests is usually charged to the package-to-board expansion mismatch, but the strain is concentrated by the package's warpage, which is set upstream by the moulding and the die attach. A change in the die attach can therefore appear as a solder failure with no change to the solder. Testing two die-attach variants in the same package and board would move joint life without touching the joint.
Inherited expansion
H4374. The strain range in a solder joint includes the board's own expansion, and for small packages that term can exceed the package's, so the joint measures the board's properties as much as the package's and a board material change can dominate a package change. Two board materials with one package give the term directly.
Geometric transition
H4375. Below a standoff height a joint stops behaving as a compliant bridge and starts behaving as a constrained layer, and the crack path moves from the bulk to the interface. The transition is geometric, set by the ratio of standoff to joint width, and it explains why thin joints fail differently rather than merely sooner. A standoff series read with fracture paths places the transition.
Roughness inversion
H4376. A thick intermetallic layer is normally read as the precursor to a brittle joint, but a thin continuous layer can be worse than a thick scalloped one, because scallops deflect a crack while a flat interface concentrates it. The relevant variable is the interface's roughness rather than the layer's thickness, so two joints of equal average thickness and different morphology should rank against the thickness rule.
Continuum claim
H4377. Head-in-pillow and a plain open joint are the same defect at two stages, distinguished only by whether the ball deformed before the paste's flux was exhausted. Treating them as separate modes has produced two sets of process fixes for one oxide problem. Interrupt a reflow series at different times and the continuum appears.
Destructive artefact
H4378. Dye-and-pry creates the cracks it reports, because the dye enters only where a gap already exists and the prying then extends it, so the fracture surface belongs to the test and not to service. The method still finds where a gap is, and it cannot measure how large the gap was. Set dye results beside a non-destructive image of the same joint and the artefact is obvious.
Anisotropy claim
H4379. The stiffness of a solder joint depends on the orientation of its beta-tin grains, so the same joint can be strong along one board axis and weak along the other. Grain orientation follows the thermal gradient during solidification, and that gradient follows the layout around the joint. Two joints of identical geometry with different neighbours differ in stiffness by an amount that follows their grain maps.
Test-dependent ranking
H4380. Adding bismuth to a tin-silver-copper alloy improves thermal fatigue and worsens drop resistance, so the ranking of alloys reverses between the two tests and no alloy is best at both. The same solid-solution strengthening that helps under slow strain embrittles the joint under fast strain. One alloy series through both tests crosses at the alloy that is second-best at each.
Path crossover
H4381. At fine pitch a joint's life is limited by consumption of the pad, and at coarse pitch by creep of the bulk, with the crossover where the intermetallic thickness becomes a set fraction of the standoff. Below the crossover, changing the alloy does little, and above it, changing the pad finish does little. A pitch series with both measurements crosses where the limiting path changes.
Parameter count
H4382. A solder joint under thermal cycling has two independent strain components, shear and peel, so a life model with more than two independent parameters is fitting the test rather than the joint. The number of independent parameters is a property of the geometry and not of the alloy. A deliberately asymmetric pad separates the two components.
Location bias
H4383. The joints that fail first in a thermal cycling test are those nearest the die corner, but the first to fail in service can be at the package edge, because the board's constraint differs from the fixture's. Selecting the corner joint for analysis therefore studies the fixture. Compare failure locations on a free board and a clamped board and the shift is the fixture's.
Position-dependent risk
H4384. A void at a solder interface is harmless until the joint creeps enough to turn it into a crack, so its danger is set by its position relative to the maximum-strain path rather than by its area. Two voids of equal size in different places therefore have different consequences, which is why void-fraction limits predict so little; a controlled void at two positions settles it.
Rate criterion
H4385. The criterion for solderability should be the rate at which the wetting force rises rather than the final force, because a poor wetting system reaches a respectable final force given enough time. The rate follows the flux's activity and the pad's oxide, which are the variables that matter. A wetting-balance run reported with the slope as well as the plateau would separate good from merely slow.
Rate competition
H4386. A solder joint's life is the minimum of three rates, creep, intermetallic growth, and pad consumption, and the minimum changes with temperature, so a qualification at one temperature ranks alloys for that temperature alone. A ranking is therefore a statement about the test and not about the joint. Two temperatures that swap the limiting rate invert the ranking.
Electrostatic discharge
Location discriminator
H4387. Devices that fail at incoming inspection are often charged to electrostatic discharge, but an assembly defect can have the same signature and a different location, and location is the discriminator. A discharge concentrates at the protection device or the gate, while an assembly defect concentrates at a bond or a pad edge. Mapping the failure site against the layout separates them without any electrical test.
Fixture term
H4388. The peak current in a charged-device discharge includes the tester's own capacitance and the socket's, and at small package sizes those terms can be comparable to the device's. A passing level is therefore partly a property of the handler, which is why the same part can pass on one line and fail on another. Replacing the device with a dummy of known capacitance gives the fixture's share.
Lumped-to-distributed transition
H4389. There is a package size above which a charged-device discharge stops being a lumped event and becomes a distributed one, so the protection clamp sees only part of the current and triggers late. The threshold follows the package's internal inductance relative to the discharge's rise time. Two package sizes with the same die bracket the transition.
Size paradox
H4390. A larger protection diode has lower on-resistance and more capacitance, and past a size the capacitance slows the clamp enough that the let-through energy rises, so a smaller diode can protect a fast transient better. The optimum follows the transient's rise time rather than the diode's steady rating. Two diode sizes at two rise times cross at the optimum.
Analysis-created damage
H4391. The melt filament that failure analysis reports as discharge damage can be created by the analysis itself, by the focused ion beam's current or by the curve tracer used to find the short. The signature of a created filament is that it lies under the probe or inside the milled region. Imaging before any electrical probing, on a neighbouring untested device, would establish what the process leaves behind.
Energy criterion
H4392. The right criterion for discharge robustness is the energy delivered at the device's own failure threshold and not the passing voltage, because two devices with different current-voltage shapes can pass the same voltage with different energies. Voltage is a convenient number and a poor ranking variable, so a test that records the current waveform as well as the voltage would reorder a supplier's catalogue.
Moisture, pressure and the sealed volume
Mechanism swap
H4393. Delamination after moisture preconditioning is usually explained by vapour pressure opening a gap, but the moisture's larger effect may be to lower the interface's fracture energy, so a gap opens under stresses that were always present. The two explanations differ over whether a vented package would still delaminate, because pressure needs a sealed volume while embrittlement does not. A package with a deliberate vent, soaked identically, would decide it.
Cleaning term
H4394. The moisture absorbed during aqueous cleaning after die attach is larger than the bake that follows removes, so the package entering moulding is not dry and the preconditioning starts from an unknown level. The term is invisible because the bake is specified by time and temperature rather than by the part's actual content. Weighing the same part after cleaning and after the bake gives the term.
Perimeter threshold
H4395. There is a die perimeter above which a popcorn crack starts at the die corner regardless of the compound, because the corner is where the vapour's path to the free surface is longest and the constraint is highest. The threshold is geometric and moves with the die's aspect ratio. A die-size series at fixed thickness marks the onset.
Bake penalty
H4396. Baking a package before assembly normally improves moisture resistance, but over-baking can worsen it by oxidising the leadframe and lowering the adhesion that the moisture test later exploits. The penalty begins where the oxide's growth outruns the water's removal. A bake-time series read with adhesion and moisture sensitivity together has a minimum.
Stage identity
H4397. Popcorn cracking and delamination are one event observed at two stages, and the difference between them is only whether the crack has reached a free surface. Because they are one process, a specification that limits delamination area while ignoring internal cracks can pass a part that is one cycle from popping. Cross-sections of parts that passed the acoustic criterion contain the internal cracks.
Failure analysis and its instruments
Preparation artefact
H4398. Decapping a package changes the stress state the analysis is trying to observe, so a die crack found after decap may have been opened by the decap rather than by service. The artefact is recognisable when the crack faces fit without strain and show no corrosion or fatigue marking. Image the same part before and after decap and the preparation's own contribution is visible.
Extreme-unit bias
H4399. A failure analysis studies the unit that failed worst, so its root cause is the cause of the extreme rather than of the population, and a corrective action built on it can leave the common cause untouched. The bias is structural, because the part is selected for having failed rather than sampled for being representative; analysing a random failing unit beside the worst one tests whether the causes agree.
Elimination standard
H4400. A failure analysis is only as good as the set of hypotheses it eliminates, and the usual root cause is the first candidate that fits rather than the one that survives the others. The discipline that would improve it is the same as in any experiment, which is to state what would be seen if the candidate were wrong and then go looking for it. A report that lists the eliminated candidates is more useful than one that lists only the survivor.
Status and plan
One hundred entries, numbered in sequence. Each was composed on its own, with the argument shape fixed before the subject, so that no two share a form; the label above each claim is there to be checked against the sentence beneath it rather than accepted as a taxonomy.
The physics stays inside the package: the attach as a compliance and a conductor at once, the wire as a chain of three joints, the compound as a constrained viscoelastic solid, the thermal path as a set of bottlenecks that move with power, the solder joint as the place where every upstream decision is spent, and the instruments that can create what they report. No claim here concerns a brand, a standard, or a person.
If one of these is worth an afternoon, it is H4309: the stress-free temperature of a die attach is usually treated as a material constant, and the claim is that it is the temperature at which the attach's relaxation time matches the cool-down time, so it moves with the cooling rate. One paste, one die, three cooling rates, and a die-curvature measurement at room temperature would either give three different apparent constants or retire the idea.
Incoprea Research · hand-considered hypotheses 4301–4400 · 3 November 2026 · ~/Desktop/Nov3/