Quality evidence and reproducible processing

A process window is evidence-dependent

An acceptable combination of settings must produce the required state consistently over relevant feedstocks, machines, positions and part geometries. An acceptable coupon is useful but does not automatically qualify every component.

Imagine a process that repeatedly makes strong coupons near the centre of its build plate, while an edge region produces weaker ones. A pooled average can look satisfactory because the stronger positions compensate numerically for the weaker ones. A component containing the weak region does not gain that compensation physically. Begin with the quantity and location the function needs, then ask which evidence can actually answer that question.

A process window is a bounded set of input conditions for which the required outputs have supporting evidence. Inputs include material identity and condition, machine state, part geometry, orientation, build location and later treatment, as well as the settings file. A nominally identical setting need not give an identical local thermal history or final state. Changes can alter heat removal, consolidation or subsequent structural evolution. These mechanisms explain why records matter; they do not identify the cause of every observed difference.

Repeatability in measurement concerns precision under specified repeated measurement conditions, usually over a short period with the same procedure, operator, system, location and operating conditions. Broader measurement reproducibility deliberately includes changed conditions such as location, operator or system. Manufacturing also needs a stated set of conditions: repeated coupons in one build probe a narrower question than builds made on different machines. Do not treat a metrological vocabulary definition as a universal manufacturing qualification rule.

Distinguish the object being repeated. Four nondestructive readings of one coupon can probe measurement scatter; three separately manufactured coupons also contain manufacturing variation. A destructive tensile test cannot be repeated on the same intact coupon after failure. Coupons sharing one build also share material and thermal events, so their count is not a count of independent builds. Good agreement under one set of conditions says little about an untested change.

Keep position and build identity visible

The following data are entirely invented. Three separate tensile coupons are tested at each of two labelled build-plate locations in each build. All have the same coupon geometry, tensile orientation, final preparation and test procedure. The quantity of interest is yield stress S in MPa under that procedure; the table reports estimates of S. Each sampled coupon’s S must meet the invented screen S ≥ 500 MPa. This is a teaching screen, not a material specification, component allowable or sufficient qualification test.

Builds A and B use the same recorded machine and feedstock lot on separate runs. Build C uses both a different machine and a different feedstock lot, with the same nominal settings. The centre and edge labels designate sampled locations, not every point in a component. The first comparison treats the supplied estimates as exact; measurement uncertainty is added separately below.

Every cell is a tight cluster, yet the cell means are not interchangeable. The centre means are all 520 MPa. The edge means are 506, 504 and 493 MPa. The observed ranges within cells are only 2–4 MPa. These small ranges do not remove the lower edge results, and they do not estimate a stable long-term distribution from three coupons.

Pooling all 18 values gives 510.5 MPa. That is a correct arithmetic mean and the wrong substitute for an each-coupon screen: all three C-edge values fail. Equal centre results also do not establish that C reproduces the whole sampled spatial pattern. Machine and feedstock changed together, so the poorer edge cannot be assigned uniquely to either. Retain the failed cell instead of deleting it as inconvenient.

A decision needs an explicit measurement rule

Now separately stipulate that a defensible analysis bounds each coupon’s actual value within ±3 MPa of its estimate. This is a guaranteed interval only inside this invented model. It is not a standard or expanded uncertainty, not a confidence interval estimated from the three coupons, and not reduced by averaging different specimens. A real uncertainty report needs its own meaning and coverage justification.

For this exercise use a conservative interval rule: a coupon establishes the lower screen only if its entire interval lies at or above 500 MPa; it establishes failure only if its entire interval is below 500 MPa. Otherwise the result is unresolved under this rule. If y is the estimate and b the bound, the tests are y − b ≥ 500 for established pass and y + b < 500 for established failure. Equality at the lower limit passes because the screen is inclusive.

With b = 3 MPa, the smallest A/B estimate is 503 MPa, with interval 500–506 MPa. Every sampled A/B coupon therefore establishes this particular screen. The highest C-edge estimate is 494 MPa, with interval 491–497 MPa; all C-edge intervals remain below the limit. A changed estimate of 501 MPa would instead have interval 498–504 MPa: neither passing nor failing is established. Rounding it to a positive margin would hide the unresolved decision.

These are results for sampled coupons under a stipulated property rule. They do not prove every unsampled region, another tensile direction, fatigue response, a larger geometry or later service conditions. Dimensional, structural and response requirements can be independent; dense material is not automatically crack-free or strong enough. An instrument check does not by itself establish representative sampling.

A negative inspection has a conditional meaning

Nondestructive testing examines a part without deliberately destroying it, but a flaw must generate an observable signal in the inspected arrangement. Shape, orientation, material path, location, access, noise and the decision threshold can change detection. A demonstrated response to one flaw family does not establish detection of every harmful condition. A process-monitoring signal also needs a validated relation to final defects or properties before it can replace their examination.

Consider a separate invented probability model for one precisely specified harmful flaw class and inspection arrangement. A part contains either one such flaw or none. Among parts with that flaw, the chance of a positive indication is 0.70; among parts without it, the chance of a positive indication is 0.02. These probabilities are supplied, not estimated from the coupon table, voxel size or a few seeded trials. They are not characteristics of a named instrument.

Suppose the flaw prevalence in the inspected population is separately known to be 0.01. Out of 10,000 parts, expected counts are 100 flawed and 9,900 unflawed. Expect 70 positives and 30 negatives among the flawed parts, and 198 positives and 9,702 negatives among unflawed parts. Expected counts describe averages across populations in this model, not guaranteed observed counts in a batch.

There are then 9,732 expected negative indications, including 30 flawed parts. The conditional probability of the specified flaw given a negative result is 30/9,732 ≈ 0.003083, or 0.3083%. The 30% miss probability instead conditions on a flaw already being present. Confusing those two denominators changes the question. If prevalence is unknown, these supplied detection/false-positive probabilities alone cannot determine the probability for a negative part.

Three inspections with the same 0.70 detection probability would miss a present flaw with probability 0.30³ = 0.027 only if their outcomes were conditionally independent for that flaw and setup. Reusing the same view can repeat the same blind spot. More scans are not three independent opportunities merely because three files were saved. The remaining probability is also about this flaw class, not every reason a component can fail.

Trace a changed process claim to discriminating evidence

Traceability here means keeping a recoverable relationship between a part and its material, build, location, geometry/orientation, machine state, software/settings revision, later treatment and inspection/test records. It is distinct from metrological traceability of a measurement result to a reference through a documented calibration chain. Both can matter; one does not supply the other.

For the C-edge discrepancy, preserve the original data and check the measurement arrangement before assigning a processing mechanism. A useful next comparison could cross the two machines with both feedstock lots, keeping other chosen conditions controlled and including multiple separate builds and the relevant positions. Randomising or balancing run order helps avoid confusing a machine difference with time drift. Such a design can discriminate effects better than simply repeating only the favourable centre coupons. It still needs enough evidence for the intended claim and does not by itself solve every interaction.

A primary interlaboratory study of additively manufactured cobalt-chromium reported larger differences between laboratories than within laboratories using nominally identical laser systems. This particular observation shows why the conditions and sampling hierarchy deserve attention; it does not imply a universal variability ratio or establish which uncontrolled variable caused a particular difference. The numerical coupon and inspection cases here are separate original models.

If a numerical process model fits the same runs used to adjust its parameters, that fit is calibration evidence. Testing a specified prediction against independent relevant observations addresses a different validation question. Compare like quantities and account for measurement and model uncertainty; matching a surface signal is not automatically validation of internal stress, hidden porosity or lifetime.

The defensible conclusion for the supplied coupon dossier is narrow: A/B meet the stated screen at the sampled positions under the supplied bound, while C-edge does not. The data identify where a broader reproducibility claim fails; they do not authorise every part from any of these builds. Record what changed, what remains untested and which new evidence would change the conclusion.

Inspect what the function requires

Use dimensional, structural and property checks with known detection capability. Define acceptance around function and relevant failure modes. Preserve traceable process records so a later change can be assessed rather than treated as an identical repeat.

Worked examples

Try the reasoning

1. Why retain feedstock identity?

   Show the answer and reasoning

   Composition, contamination and size distribution can affect processing.

   The input is part of the process state.

2. Does nondestructive inspection detect every flaw?

   Show the answer and reasoning

   No.

   Detection capability depends on size, shape, orientation and method.

3. Keep the coupon table but change the supplied bound from ±3 to ±5 MPa. Does the B-edge group establish that every one of its three coupons meets 500 MPa?

   Show the answer and reasoning

   No. 503±5 spans 498–508 and 504±5 spans 499–509; those two are unresolved. 505±5 spans 500–510 and establishes the inclusive screen.

   The nominal values have not changed, but the evidence rule uses each interval’s lower endpoint. Failure is not established for the first two because their upper endpoints also exceed 500.

4. With the original ±3 MPa bound, raise only the invented lower screen to 505 MPa. Does A-edge establish a pass for every coupon?

   Show the answer and reasoning

   No. Its 506 and 504 estimates have lower endpoints 503 and 501 MPa. Only 508 has lower endpoint 505 MPa and establishes that screen.

   Changing the requirement changes the decision. The first two intervals cross 505, so they are unresolved under this rule, rather than proven failures.

5. One nondestructively tested coupon gives 503,504,503,504 MPa-equivalent instrument indications under repeated placement. Is this four independent manufactured specimens or four independent builds?

   Show the answer and reasoning

   Neither. It is one specimen with four measurement repetitions.

   The repetitions may inform that measurement procedure’s scatter. They do not sample specimen-to-specimen variation, spatial process variation or separate build histories. A conversion to a material property would also need its own validated relationship.

6. Change only the specified-flaw prevalence in the inspection model from 1% to 10%. For 10,000 parts, what fraction of negative indications are now flawed?

   Show the answer and reasoning

   There are 300 flawed negatives and 8820 unflawed negatives, so 300/9120≈3.2895%.

   1000 flawed parts yield 0.30×1000=300 misses; 9000 unflawed parts yield 0.98×9000=8820 negatives. The same 70% detection probability leads to a different negative-result probability when prevalence changes.

7. A flaw is hidden behind the same obstructing feature in three identical views. Can you automatically claim a miss probability 0.30³?

   Show the answer and reasoning

   No. The supplied single-inspection probability does not establish conditional independence of the repeated views.

   A common blind spot can make outcomes strongly correlated. At the perfectly correlated limit three repetitions can have the same 0.30 miss probability as one. The actual dependence and relevant flaw/geometry family need evidence.

8. A report calls 90 coupons from three builds “90 independent demonstrations across builds”. What is wrong, and what useful information can still remain?

   Show the answer and reasoning

   There are only three separate build histories. Coupons within a build may still reveal specimen and position variation, but they share build-level influences.

   Keep the hierarchy and labels. More coupons can improve knowledge within the sampled builds; they do not manufacture 87 additional independent build histories or establish reproducibility on another machine.

What to carry forward

• Qualify the relevant process–part combination.
• Use traceable records and appropriate detection limits.
• Keep measured repeats, specimens, positions and builds distinct; retain failed groups instead of hiding them in a pooled mean.
• Compare the actual requirement with the measured quantity and a stated uncertainty/decision rule.
• Detection given a flaw and flaw probability given a negative result have different denominators.
• Extend a process claim with relevant changed-condition evidence, recoverable records and tests that can separate competing explanations.