Small Bets in Thermal Interface Optimization That Compound

You're staring at a sintering curve on the monitor. The line dips, then climbs, then flattens—textbook, maybe. But what you really need to know is the porosity inside the joint, and that curve won't tell you directly. Not unless you know how to read it. This isn't about waiting for the thermocouple to move; it's about extracting the story before the process even finishes.

Porosity is the silent killer of thermal performance. Too much voiding, and your heat path turns into a bottleneck. But measuring it after the fact—cross-sectioning, acoustic imaging—costs time and parts. So engineers look for early signals. Sintering curves carry those signals, if you know where to look. This guide walks you through the decision: who must choose, what options exist, how to compare them, and how to implement a reading method that doesn't wait for the thermocouple.

Who Decides Porosity Targets and When?

Role of process engineers versus thermal designers

Porosity targets rarely come from the thermal designer’s desk. They come from process owners who have already watched one sinter run fail shear testing at hour forty. The thermal team cares about junction temperature, sure—but the process engineer owns the press profile, the paste batch, and the screaming match with the substrate supplier when voids show up on X-ray. That ownership forces early decisions. Wrong order—thermal designer sets a porosity number, process engineer nods, then the real work begins. The catch is that porosity behaves like a stubborn dependent variable. You can't just dial it in. Sinter pressure, ramp rate, and peak temperature all interact, and each interaction shifts the void distribution in ways that only show up on cross-sections days later.

So who actually signs off on the target? In my experience, the process engineer proposes, the thermal designer disposes—but only if the number makes physical sense. A target of 8% porosity sounds clean on paper. That sounds fine until you realize your paste formulation struggles below 12% without cracking the die. Then the deadline pressure kicks in. You have a prototype gate in three weeks, and someone upstairs assumed porosity is a simple spec, like a resistor value. It's not.

Timeline: before first prototype or after yield drop?

Set the target before the first prototype runs. Not after first yield drop, not during qualification review, not when the customer’s thermal simulation shows a hot spot you can't explain. I have seen teams skip this step because they were busy chasing deposition thickness. Then the first sinter batch comes out at 14% porosity, nobody knows if that matters, and the schedule eats three weeks of cross-section analysis to answer a question that should have been answered in a half-day meeting. The timeline problem is real: if you wait until yield drops, you're already reacting to a symptom. Porosity targets set late become excuses, not specifications.

But there is a subtler trap here. The process engineer might set a target that's too tight because they want a safety margin. That hurts yield for no thermal benefit. A looser target, say 15% instead of 10%, might cut cycle time by twenty percent because you can use a faster ramp. We fixed one such case by testing three porosity levels and measuring actual Rth, then specifying the upper bound that still met thermal budget. Nobody had bothered to ask whether the tight spec was technically necessary. The trade-off is between over-engineering and under-specifying, and the decision window is narrow.

Porosity targets are commitments made before data exists, then defended with data you're too busy to collect.

— process engineering lead, power module line

Claim desks that separate intake verbs from appeal verbs stop copy-paste denials from looking like thoughtful casework under audit lights.

Contract manufacturer’s stake in porosity specs

If you outsource sintering, the contract manufacturer has an even bigger stake. They quote cycle time, tool wear, and scrap rate based on your porosity spec. Tighten the spec after qualification, and their margin disappears. That's when they start pushing back with batch-to-batch paste variability arguments—some legitimate, some not. The smart play is to agree on the porosity target and the reading method before signing the contract. Name the cross-section locations, the sampling frequency, and the acceptance criterion. Don't leave it vague. A vague spec means the CM’s definition of "acceptable porosity" becomes the one that passes their internal QC, which may not match your thermal needs.

What usually breaks first is communication. The CM sees a curve shift on their sinter press and doesn't mention it. You see a thermal failure in the field and blame their process. Both sides lose. The deadline pressure compounds this—once production starts, nobody wants to pause for a porosity review. That sounds like an operational problem, but it's actually a decision-making problem. The target was set late, or not at all, and now the spec is whatever the press produced. So: when do you decide porosity targets? Before the first prototype, in writing, with a stated method for verifying them. Otherwise you're not reading curves early—you're reading them after the failure, which is the most expensive way to learn.

Options for Reading Porosity from Sinter Curves

In-situ Resistance or Conductivity Tracking

Watch the electrical path as the die-attach layer densifies. Silver or copper sinter pastes drop resistance sharply once particle-to-particle contact forms a continuous network. That transition happens early—often before the curve looks visually complete. Attach thin leads to the substrate or use the die itself as one electrode. The signal is noisy at first, then stabilizes. I have seen teams mistake that stabilization for full densification, only to find porosity at 18% later. The resistance floor is not the finish line; it's the point where the network percolates. Use it as a trigger for further verification, not as a verdict.

Vendor reps rarely volunteer the maintenance interval; however boring it sounds, the calibration log is what keeps tolerance from drifting into customer returns.

One pitfall: contact resistance from probe placement can swamp the bulk signal. Fix the probe geometry once and don't vary it across lots. Relative shifts matter more than absolute values—track the slope of resistance decay, not the final number.

Curve Shape Parameters: Slope, Plateau, Time-to-90%

Every sinter profile leaves a fingerprint in the pressure or displacement curve. Plot the tool position against time and look for three landmarks: the initial compaction knee, the mid-hold plateau, and the final creep. The slope between knee and plateau correlates with how fast pores collapse. A shallow slope means slow pore closure—expect higher residual porosity. The plateau length tells another story: a flat, long plateau often hides trapped gas that can't escape. Time-to-90% of final displacement is a simple metric, but it captures process consistency better than most operators think. Track it per part, not per batch.

The catch is that curve shape responds to paste batch viscosity and substrate flatness before it responds to porosity. Normalize the raw curve against a reference part from the same paste lot. Otherwise you will chase noise.

Ultrasound or Acoustic Microscopy Off-Line

Pull a part, scan it acoustically, and map the reflection amplitude across the bondline. Dense sintered silver returns a uniform, low-reflection pattern; porous regions scatter and show bright spots. This is not inline, but it's faster than cross-sectioning and gives a two-dimensional porosity map. Use it to calibrate the curve parameters—run three parts at low, medium, and high pressure settings, scan them, then correlate the acoustic signature to the slope values. That calibration holds for weeks if the paste stays consistent.

Kill the silent step.

However, acoustic scans miss sub-micron closed pores. Cross-sections still rule for absolute numbers. The trade-off is speed versus resolution, and for production monitoring, speed usually wins.

Process Signature via Pressure and Dwell Time

Press force curves during the hold phase often reveal pore collapse dynamics better than displacement. A pressure sensor under the substrate shows micro-relaxation events—tiny drops in force as pores implode. Count those events or integrate the force ripple over the dwell window. More ripple usually means more pores collapsing late, which points to porosity that stays open too long. Some process engineers plot the derivative of force against time; a smooth, decaying derivative suggests uniform densification, while spikes indicate localized void collapse.

'The curve is not porosity. It's a shadow of porosity—useful only if you know the light source.'

— process engineer reflecting on calibration limits

That said, pressure signatures drift with tool wear and backing film compliance. Recalibrate after every 500 parts or whenever the baseline shifts. The real question is not which method to use—it's which one your team will actually review before the lot ships. Pick one curve feature, plot it on a control chart, and flag deviations early. That beats any fancy model that nobody reads. Start with the slope and time-to-90%; add acoustic checks on first articles and after any paste change.

Field note: gaming plans crack at handoff.

Puffin driftwood stays damp.

Field note: gaming plans crack at handoff.

Trail guides who log bailout routes before summit weather windows treat courage as a checklist item, not a brand slogan on new gear.

Criteria That Separate Useful Porosity Signals from Noise

Sampling rate and sensor placement

A porosity signal is only as good as the data feeding it. If your thermocouple sits 15 mm away from the sinter layer, you're reading the die attach’s mood, not its state. Watch the response time of the sensor against the ramp rate of the press. A slow sensor on a fast ramp blurs the inflection points where void collapse actually shows up. Place the sensing element as close to the sinter interface as the tooling allows—inside the die, under the substrate, or in the punch itself. And sample at least five times faster than the fastest thermal event you care about. Otherwise, you catch the echo, not the moment.

Sampling rate costs memory and processing time. More data is not automatically better. We once ran a trial at 100 Hz for a week straight, and the file was so bloated we could barely plot it. The useful window was about four seconds per cycle. Fixed that by triggering recording only around the squeeze phase. That cut storage by 90 percent and made the curves legible again.

Sensitivity to void morphology vs. total void fraction

Two parts can show the same total porosity and behave completely differently in thermal cycling. A single large void near the die corner will crack faster than ten tiny pores spread evenly. Total void fraction tells you how much air is trapped. Morphology tells you where the danger lives. The catch is that curve-based methods—pressure decay, thermal diffusivity, or ram displacement—mostly respond to total volume, not shape. They see the forest, not the snags.

So what separates a usable signal from a decorative one? The useful method is the one that tracks the failure mode you actually care about. If your product dies from corner delamination, a total-fraction reading will miss it every time. Cross-sections give you morphology, but they kill the part. Inline curve methods give you total fraction, live and cheap. Neither is enough alone.

That's why the correlation test matters more than the raw measurement precision.

Correlation strength with post-process thermal impedance

Don't validate a porosity-reading method against porosity itself. Validate it against the thing your customer feels: thermal impedance rise after 500 cycles. I have seen engineers polish a curve-fit until it predicted void fraction with 98 percent accuracy—and then discover the whole exercise was useless, because void fraction only explained 40 percent of the impedance drift. The signal that matters is the one that leads the failure, not the one that matches the lab report.

Run a designed experiment. Vary pressure, temperature, and paste batch. Measure your curve signal, then measure thermal impedance on the same parts. Check the R² against impedance, not against destructive porosity readings. If the curve correlates with impedance above 0.7, you have a watchdog. Below that, you have a thermometer with an opinion.

That order fails fast.

Cost per part and cycle time impact

Here is where good ideas go to die. A method that adds eight seconds per cycle on a 30-second process is a 27 percent throughput hit. That's not a measurement, that's a tax. Inline optical or ultrasonic systems can push into six figures per line. Pressure-decay sensing, if you already have a closed die, might cost only the plumbing and a transducer. Same information class, wildly different price tags.

Claim desks that separate intake verbs from appeal verbs stop copy-paste denials from looking like thoughtful casework under audit lights.

The trade-off is brutal but simple: if the method protects against a failure that costs $40,000 in rework per month, then a $2,000 sensor that slows the line by 2 percent is a steal. If your defect rate is one part in ten thousand, skip the fancy rig and do a weekly cross-section. Wrong order—buying capability before knowing the defect cost—is how most porosity monitoring budgets get wasted.

“A porosity signal that costs more than the defect it prevents is not quality control. It's theater.”

— process engineer, power module line, after scrapping a $120k inline scanner

Start with a cheap sensor and a spreadsheet. Prove the correlation to impedance. Then spend on the permanent installation. That sequence has never failed me. The reverse has failed plenty.

Trade-offs: Destructive vs. Inline Porosity Checks

Cross-sectioning: gold standard but kills the part

Destructive porosity checks give you truth. Polish, etch, image, count pores—you see exactly what the sinter layer did. No inference, no math. But the part is gone. For a power module that costs $200 in materials and another $400 in processing, that stings. I have watched teams grind down ten modules a week just to confirm what they already suspected from two blistered samples. The data is beautiful. The scrap bin fills fast.

The catch is that cross-sections only answer the question you asked at that moment, at that spot. A single slice is a 2D line through a 3D pore network. You might hit a dense patch or a void cluster. The sampling error is real. Cutting more sections helps, but each cut costs time and money. This is why destructive checks work best as periodic audits, not as a continuous hygiene monitor.

A mentor explained that however polished the dashboard looks, the pitfall is skipping the failure rehearsal that would have caught the silent assumption on day one.

Acoustic microscopy: non-destructive but resolution limits

Scanning acoustic microscopy (SAM) solves the “kill the part” problem. You scan, you image, you rework or scrap based on the void map. Heavy automotive modules often get 100% SAM screening—it pays off when a single field failure triggers a recall. That said, SAM has blind spots. Resolution dips when pores shrink below roughly 30–50 micrometers depending on frequency. Fine porosity—exactly what affects thermal impedance at die attach thickness—blurs into speckle noise.

Most teams skip this: SAM also struggles with buried interfaces near edges or under thick copper. Good signals require careful coupling, flat surfaces, and operator skill. Machines run from $80k to $300k. For high-mix labs, the setup time per recipe eats the throughput advantage. Quick reality check—you gain non-destruction but lose speed, and you still can't read temperature behavior directly from the image.

Curve-derived porosity: fast but indirect

Sintering curves—pressure, temperature, displacement, or electrical resistance over time—offer a third path. Density and pore structure leave fingerprints in how much the die settles, how fast the paste collapses, how resistance drops. You read the curve, estimate porosity, and flag deviations. No killing, no expensive scanner. Just sensors you already have.

But it's indirect. The curve integrates many effects: paste batch viscosity, surface oxide, tool wear, temperature gradients. A porosity estimate from displacement alone can drift by 5–8% absolute if the paste lot changes. You trust it only after you have correlated it against cross-sections for each new material set. That correlation effort is real work. I have seen engineers burn two weeks building the lookup table—then save months of scrapped parts later.

Name the bottleneck aloud.

Table comparing methods on speed, cost, accuracy

Here is the straight comparison, no fluff:

Method	Speed	Cost per part	Accuracy	Part survives?
Cross-section	20–40 min	$50–150 + labor	High (2D slice)	No
Acoustic microscopy	2–5 min	$5–20 (amortized)	Medium (coarse pores)	Yes
Curve-derived	<1 min	<$1 (sensor wear)	Medium (needs correlation)	Yes

The accuracy column hides the real trade-off. Cross-section wins on truth, loses on throughput. SAM wins on part survival, but you can't resolve fine pores. Curves win on speed and cost, but every estimate carries a calibration debt. The best porosity watch doesn't pick one—it blends all three: curves for every part, SAM for spot checks, and cross-sections when the curve signal goes weird.

It adds up fast.

What usually breaks first is the calibration. Teams launch with curve-derived porosity, skip the cross-section correlation, then wonder why yield drops. Fix that early: run 20 destructive sections across five lots, map curve features to measured porosity, and lock that correlation as a living document. Update it when paste supplier or sinter profile changes. Wrong order means you trust a slope that lies.

So don't treat these as rivals. Treat them as a funnel: curves flag, SAM confirms, cross-section explains. That sequence costs less than weekly destructive checks and catches more than SAM alone. Start there.

Implementing a Curve-Based Porosity Watch: Steps

Calibrating baseline curves with known porosity samples

Before you trust a single curve, you need ground truth. Take five to seven sintered parts—some deliberately under-sintered, some over-sintered, a few right in the sweet spot—and measure their actual porosity with cross-sections or Archimedes density. Then overlay their recorded sinter curves. The pattern jumps out fast: peak press displacement, hold-phase slope, and the plateau height at the end all shift with porosity. Map those features to measured values and you have a local translation table. It won't be universal. Different paste batches shift the baseline. That's fine. You're building a reference for your line, not a physics textbook.

The calibration needs repeating whenever paste lot changes or the substrate finish changes. I have seen teams skip this and chase phantom porosity alarms for a week—turns out the new paste simply sinters faster. Wrong order here is costly: curve-fitting without measured anchors is astrology with a thermocouple. Fix the reference first.

Setting alarm thresholds and control limits

Thresholds come from your calibrated baseline plus process history. Pull the last thirty good parts—ones that passed cross-section checks—and compute the mean and standard deviation for each curve feature you track. Set warning limits at two sigma, action limits at three. That sounds standard, and it's. But the catch is which features you watch. Peak displacement alone will fool you; it responds to paste volume variation as much as porosity. Combine it with the hold-phase slope ratio, and false alarms drop sharply.

Use a two-out-of-three rule for warnings: two consecutive parts drifting past the warning limit triggers review, not just one blip. One-off spikes are often substrate roughness or a slight die tilt. Two in a row means something settled in. The trade-off is sensitivity versus disruption—too tight and operators ignore alarms by lunchtime, too loose and you catch failures after reflow, not before.

Integrating with existing data acquisition and SPC

Most sinter presses already log force, temperature, and displacement every hundred milliseconds. The data is sitting there. Pull it into your existing SPC software rather than building a parallel system—parallel systems die quietly when someone forgets to update them. Map the curve features to a simple attribute: pass, watch, fail. That keeps the operator screen clean. No need for full curve visualization on the floor; the trend chart of the extracted parameter is enough.

According to field notes from working teams, the boring baseline check prevents more failures than a brand-new framework introduced mid-sprint under pressure.

When the same sentence length repeats for a whole chapter, readers feel the template even if every claim is true, so break the rhythm on purpose.

What usually breaks first is the time stamp. If press data and inspection data live on different clocks, matching them becomes guesswork. Fix that before wiring any alarms. A minute of mismatch on a 20-second sinter cycle makes the correlation meaningless.

Training operators to spot anomalies

Operators don't need to understand viscoelastic models. They need three visuals: a normal curve shape, a curve that drifts slowly, and a curve that snaps off early. Show them side-by-side examples from your own line. Explain that a flattened hold phase usually means the paste didn't densify, while an early plateau often indicates premature closure of surface pores. That distinction matters for the fix—different paste, different pressure profile, different temperature ramp.

Most teams skip this step and rely on automatic alarms. Then the alarm goes off at 2 a.m., nobody knows what to check, and they reset the system to keep production moving. That hurts more than no alarm at all.

Teach operators what a good curve looks like blindfolded. The machine alarms after the fact; human eyes catch the drift before it crosses the line.

— process engineer, power module line

Reinforce with a monthly five-minute review of rejected curves from the previous month. Ask operators what they would have done differently. That routine surfaces subtle pattern changes faster than any algorithm update. Start next week with calibration parts; set the limits by the following Monday. You will have a working watch within a month, and the cross-section lab becomes a verification tool instead of a gatekeeper.

When You Skip Early Porosity Reading, These Risks Bite

Yield loss at final test — the expensive way to learn

The most brutal outcome is also the most common: the part sinters fine, passes every inline check, then dies at final electrical test. You strip the module, find cracks radiating from the die edge, and the porosity reading you skipped three days ago would have shown you exactly why. Void content above 15% in the sinter layer doesn't fail during sintering — it fails when the substrate bends under thermal stress. That bend happens during reflow, during burn-in, during the first power cycle. By the time the tester flags it, you've already paid for the substrate, the die, the sintering step, the wire bonding, and the encapsulation. All of that cost walks out the door with the scrap bin.

Heddle selvedge weft drifts.

I have watched teams chase yield drops for a week before someone thinks to check the sinter curves. The curves were there the whole time — dip in the pressure plateau, slower densification onset, early plateau flattening. Nothing in the equipment alarmed. The process window looked normal. But the porosity signal was hiding in the curve shape, and nobody was reading it.

Field failures due to thermal cycling — the slow burn

Worse than test yield is the field return. The module works for 500 cycles, then the thermal interface degrades silently — the sinter layer delaminates, hot spots form, and the device derates or dies. Thermal cycling amplifies porosity because each void acts as a stress concentrator. One void becomes a crack; the crack propagates along the die attach interface; the crack becomes a delamination. Nobody caught it because porosity was never on the monitoring list.

That hurts more than cost. That hurts reputation.

Don't rush past.

Porosity is not a cosmetic defect — it's a latent thermal failure waiting for the right number of cycles.

— field failure analysis, power module returned after 14 months

Undetected process drift from paste batch changes

Here is the sneaky one. Your paste supplier changes a solvent ratio — slight, nothing they'd flag. The new paste sinters a bit slower, densifies a bit differently, leaves 3% more porosity. If you only check porosity destructively once a week, you might catch it after a week of bad parts. If you check cross-sections once a month, you've built a month of inventory that will fail early in life. The curve-based signal catches this on the first part — the densification rate shifts, the plateau moves, the signal tells you the paste behavior changed. The catch is you have to actually look at the curves, not just the end-of-run summary.

Wasted cost on rework or scrap

Rework on sintered die attach is rarely worth it. You can't remove and re-sinter a die without damaging the backside metallization. So the choice is scrap or ship a known-faulty part. Neither is good. The real waste, though, is not the parts you scrap — it's the production time you spend troubleshooting what a curve reading would have identified in minutes. A quick check of the porosity proxy takes thirty seconds. Cross-sections take hours. Destructive analysis takes days. The math favors early curve reading every time.

When throughput doubles without a matching documentation habit, however skilled the crew, the pitfall is invisible rework spent on heroics instead of repeatable steps.

Most teams skip this because they think porosity is a quality issue, not a process control issue. That's wrong. Porosity is the earliest cheap signal you have that something drifted — budget for it now, or pay with field returns later. The choice is yours. Start reading curves before the next batch, and you'll catch the drift while it's still cheap.

Mini-FAQ: Porosity and Sintering Curves

Can thermal impedance alone tell porosity?

Partly—but not the way most people hope. Thermal impedance reacts to pore shape, pore location, and pore connectivity all at once. A 12% porous layer with fine, evenly distributed pores can measure nearly the same impedance as an 8% layer with a few large voids clustered near the die edge. The curve gives you a blended number; porosity is a structural fact. Use impedance as a gate, not a verdict. If the gate looks suspicious, that's when you reach for cross-sections.

How often should I recalibrate my curve model?

Every time your paste lot changes. I have seen teams calibrate against one batch, then silently switch suppliers, and spend two weeks chasing phantom porosity shifts. Check after every paste lot, every new stencil design, and any change in substrate surface finish. If you run continuous production, do a quick verification sample every 40 to 60 parts. That sounds tedious until the alternative—a bad lot running for hours—hits your yield numbers. The catch is that curve models drift with tool wear too. Keep a log of calibration dates alongside your curve baselines.

What if my curve looks good but porosity is high?

Then your model is reading the wrong feature. Most likely the curve is dominated by something upstream: solvent evaporation timing, pressure ramp rate, or even the clamping sequence. The sinter curve is a composite signal. A textbook curve shape can hide poor particle packing if the process window is wide. What usually breaks first is the assumption that curve shape equals densification rate. Fix this by comparing curve shape against known-good cross-sections taken at multiple process points—not just the final output. One honest cross-section beats three weeks of curve optimism.

Porosity is a structural fact, and the curve is only a shadow of it. Verify early, verify often.

— Process engineer, power module line

Are there standards for acceptable porosity?

Yes, but they're application-specific, not universal. Automotive power modules often cite 5–15% as acceptable, while some high-reliability aerospace parts demand under 5%. The problem is that standards define measurement methods differently—some use optical microscopy on a single cross-section, others use X-ray computed tomography over the whole die. A 10% reading on one method may read 7% or 13% on another. Don't chase a number without knowing which standard your customer or internal spec refers to. That's the quickest route to a false pass or an expensive false fail.

According to field notes from working teams, the boring baseline check prevents more failures than a brand-new framework introduced mid-sprint under pressure.

Name the bottleneck aloud.

A Sober Recommendation: Start with Curve Signals, Verify with Cross-Sections

Use curve-derived porosity as a screening tool

Nobody buys a sintering press to babysit curves. Yet the same thermal profile that prints your parts also prints a story about voids—if you learn to read it. Treat that story as a first pass, not a verdict. A curve that suggests porosity above target should trigger a closer look, not a scrap bin decision. The inverse matters more: a clean curve gives you confidence to keep running, saving hours of guesswork.

I have watched teams chase phantom porosity alarms for a week, only to find a thermocouple sitting 2 mm off the die. That hurts. But the opposite failure—a curve that looks perfect while voids bloom underneath—is quieter and more expensive. The fix is calibration. Match your curve signature against cross-sections until you trust the correlation. Then use the curve to screen every part, and save destructive checks for the edge cases.

Periodic destructive verification

Cross-sections are slow, ugly, and absolutely necessary. They're also the only way to confirm what the curve implies. Schedule them like maintenance, not like punishment—every shift change, every new paste lot, every time you touch the process parameters. The sample size can be small; the cadence can't slip.

The catch is that cross-sections only tell you about the parts you destroy. That's the trade-off. You trade a few ruined samples for ground truth on thousands of surviving ones. What usually breaks first is discipline. Teams run ten sections in week one, then forget until a failure pops. Set a recurring calendar block. Automate the reminder. I have seen one stubborn engineer mark the calendar and cut one sample every Friday—six months, zero surprises.

Balance false alarms against missed defects

Too tight a threshold on curve signals and you will drown in false alarms. Too loose, and the real voids slip through. The sweet spot is boring: start with a generous band, verify what trips it, then tighten slowly. This is not glamorous work. It's the difference between a process you can defend and one you merely hope works.

“A curve is a rumor; a cross-section is a photograph. Trust the rumor enough to act, but never enough to skip the photograph.”

— me, after a week of false alarms that cost a prototype run

Start small. Pick one characteristic point on the sintering curve—the peak exotherm, the plateau slope, whatever your press shows reliably. Log it for every part. Once a week, cut a cross-section and compare. Within a month you will know whether that point means anything. If it doesn't, pick another. Wrong first guess? That's fine. The act of comparing builds the dataset you need.