Sampling, scale and extrapolation limits

A few specimens do not describe every future part

A test series samples a population. The selection method, batches and manufacturing variation determine which population it can reasonably represent. Repeated readings on one specimen and tests on multiple specimens provide different information.

A laboratory makes eight thickness readings near one corner of a metal sheet. Another laboratory measures one cut piece at each of eight locations spread across four sheets. Both reports contain eight numbers, but they answer different questions. The first mostly describes that location and reading conditions. The second begins to reveal spatial and sheet-to-sheet differences. Neither automatically represents the next year of production.

The target population is the collection about which a conclusion is wanted: perhaps this sheet, all sheets from one production run, or parts made by a stated process over a stated period. A sampled unit is an item actually selected, such as a sheet or a coupon cut from it. A reading is an observation on that unit. Keep those levels visible instead of calling every number another independent specimen.

Also define the quantity. Mean thickness, the largest local thinning and the fraction of parts below a specified limit are different questions. A good estimate of the first need not establish either of the others. Record where and when specimens were obtained, how they were prepared and what was excluded.

From a property claim to a measurement develops how to define the endpoint before choosing the observations or calculating an average.

Eight readings can cover one location or several sheets

Use the invented thickness record in the table. Each value is one measurement at a stated location, in millimetres. All specimens use the same specified measurement procedure. The two recorded locations on each sheet are not a complete scan of its surface; their mean describes these sampled locations.

The mean of the eight location values is their sum divided by eight: 8.72/8 = 1.09 mm. The sheet means are 1.00, 1.06, 1.12 and 1.18 mm. Because each sheet supplied two locations, averaging these four sheet means also gives 1.09 mm. That equality follows from the equal numbers per sheet, not from a general permission to pool any sampling plan.

Now repeat the reading only at sheet A’s left location: 0.978, 0.979, 0.979, 0.980, 0.980, 0.981, 0.981 and 0.982 mm. Their mean is 0.980 mm and their range is 0.004 mm. These repeats reveal much less variation than the eight-location range of 1.20 − 0.98 = 0.22 mm. Taking more readings at that corner would not turn it into observations of the other sheets.

The 0.22 mm range also is not a measured bound on every point or future sheet. It is the range in this finite record, which mixes location differences, sheet differences and measurement effects. A calibration correction or uncertainty assessment remains necessary; sampling more locations does not repair a common instrument error.

The first figure keeps the selected locations separate from the observations taken there. Its sheet outlines and dot positions are explanatory, not measured maps or a prescribed sampling layout. Calibration, uncertainty and repeatability develops the complementary measurement checks.

Choose observations that can answer the question

Convenient sampling can miss the variation that matters. Picking only easy-to-cut central regions may omit edges; using only one run may omit changes between runs. A narrow spread in the selected pieces can coexist with important unobserved variation. More observations from the same restricted region do not remove that selection problem.

For a question about a stated production period, identify relevant runs, positions and times before selecting pieces. One approach divides that population into meaningful groups, such as runs or edge and centre regions, then randomly selects within those groups. Random selection makes inclusion less dependent on an operator’s preferred appearance. It does not guarantee a tiny sample contains every rare condition.

Preserve the group labels during analysis. If the target is an average over manufactured parts and one run made many more parts than another, its contribution depends on that population. An equal-weight average of run means answers a different question. In the table the four sheets and two locations have deliberately equal weights; that is why the displayed average is unambiguous.

A comparison of processes needs another distinction. If treatment A is tested only on Monday’s run and treatment B only on Tuesday’s, a difference could arise from treatment, run or measurement day. Vary treatments within suitable run groups and randomise their allocation or order where feasible. The aim is to avoid making two explanations inseparable, not to erase real run variation.

There is no universal specimen count that makes a sample representative. Required evidence depends on the defined quantity, variation, selection, dependence and precision needed for the decision. More independently selected production units can answer a wider population question than many repeated readings on one unit, but neither count alone supplies a justified conclusion.

A size change can alter what fraction of material is near a surface

Even a well-sampled small specimen may not reproduce a larger component. Shape, connections, thermal history and stress distribution can change. Before assigning a new material property, identify a particular physical or geometric difference that could affect the measured endpoint.

For a simple illustration, suppose the material within 0.010 mm of each of the six faces of a cube has been altered. This is a supplied geometrical layer inside the original cube, with uniform depth δ; it is not an oxide-growth prediction. A cube of side L has volume L³. The unaltered inner cube has side L − 2δ, provided L > 2δ.

The altered volume fraction is therefore f = 1 − (L − 2δ)³/L³ = 1 − (1 − 2δ/L)³. For L = 2.00 mm, f = 1 − 0.990³ = 0.029701, or about 2.97%. For L = 20.0 mm with the same depth, f = 1 − 0.999³ = 0.002997001, or about 0.300%. A tenfold side length does not preserve the altered fraction.

This calculation predicts a volume partition for the stipulated layer, not strength, stiffness or service life. A surface flaw can govern failure even when the affected volume is small; a volume-average response may weight it differently. The same layer depth may itself fail to transfer if temperature, time, transport or chemistry changes. Specify which step is geometry and which needs new material evidence.

Units, scales and orders of magnitude supplies the length, area and volume conversion reasoning used here. The fraction above counts the combined face layers once by subtracting the inner cube, including the regions near edges and corners.

Extrapolation needs a mechanism

Using short tests to predict long life or tiny specimens to predict large parts extends beyond observed conditions. A validated physical model can help, but changing failure mechanisms can invalidate the extension. Report the observed range separately from the predicted range and include uncertainty appropriate to the inference.

Interpolation estimates a response at a value inside the investigated range of a continuous variable; extrapolation extends outside it. A temperature halfway between two tested temperatures is an interpolation in temperature. Changing from one supplier to another is not the same idea: supplier names have no numerical halfway position. Other state variables still need to match even when one coordinate lies inside its range.

Consider a second invented record made with separate matched coupons aged for different durations at one fixed environment. A tensile test destroys its coupon; these are not repeated breaking tests on a single surviving piece. Define retained strength R = 100 × σ_u(t)/σ_u(0), where σ_u is the consistently measured ultimate tensile strength and σ_u(0) is the supplied unaged reference. R is a percentage of that reference, not a probability of survival.

The reported illustrative values are R = 100, 92, 84 and 76% at t = 0, 4, 8 and 12 hours. A straight description is R_A(t) = 100 − 2t when t is inserted numerically in hours. Its slope is −2 percentage points per hour. Under that description, t = 6 h gives 88% inside the observed interval, while t = 20 h gives 60% outside it.

Here is a deliberately constructed alternative: use the same line through 12 h, then R_B(t) = 76 − 2(t − 12) − 0.25(t − 12)² beyond 12 h, with time again in hours. The quadratic coefficient has units percentage points per hour squared. This alternative agrees with every supplied observation and even has the same slope at 12 h, yet predicts 44% at 20 h. It represents an untested possibility of accelerating loss, not evidence that such acceleration actually occurs.

Good agreement over 0–12 h cannot by itself choose between these unobserved futures. A useful extension needs a mechanism expected to persist, relevant material and environmental state, and independent confirmation near the proposed use range. Neither curve is meaningful once its assumptions fail, and a formula that eventually predicts negative retained strength must not be continued literally.

In the second figure, filled markers show only the four invented observations. The solid segment is the chosen description between them; dashed segments and open endpoints are extensions. Neither a drawn line nor a predicted endpoint adds another measured coupon.

Test the proposed extension, not only the fitted line

Ask what would remain the same and what might change. Ageing could alter a phase or reaction pathway; a higher temperature could activate a different damage process; a larger piece could cool differently. A changed mechanism can matter more than the small scatter around the original fit. Faster testing is useful only when the relationship to the intended service process is supported.

In the ageing example, a next investigation could add independently prepared coupons at later times, check their structure and failure features, and retain the original environment and test definition. The observed strength endpoint and any proposed degradation indicator must actually be connected. A surface appearance that changes smoothly need not predict a mechanical failure endpoint.

Report the inference in parts: the observed specimens, positions and conditions; the model used to connect them; the extension being proposed; and the evidence or uncertainty supporting that extension. For example, “the four time points follow the stated line over 0–12 h; 60% at 20 h is its unconfirmed projection.” That lets a reader use the measured result without inheriting an unsupported lifetime claim.

Temperature, time and structural change gives a material-state example of why a thermal path cannot be reduced to a clock multiplier. Microscopy, diffraction and structure evidence develops complementary observations for testing a structural explanation. Lifetime evidence, scatter and extrapolation later treats event records and population inference; the retained-strength percentages here are a different quantity.

Worked examples

Try the reasoning

1. A test record covers 0, 4, 8 and 12 h at one temperature. Classify a prediction at 10 h and one at 18 h. Does the first classification guarantee accuracy?

   Show the answer and reasoning

   10 h is interpolation in duration; 18 h is extrapolation. Interpolation alone is not a guarantee.

   The time coordinate is inside or outside the observed range. Unmeasured structure, model shape, uncertainty and mismatched conditions can still affect either prediction.

2. An accelerated test at a much higher temperature creates a new brittle phase absent in service. Can a single time multiplier automatically convert its failures into service lifetime?

   Show the answer and reasoning

   No. The changed state can introduce a different failure mechanism.

   An acceleration relationship needs support for the mechanism and material state it connects. A convenient fit or a hotter test alone does not justify transferring that new failure route.

3. A claim concerns three production runs and both edge and centre regions. Plan A takes thirty readings at one centre location in run 1. Plan B selects pieces from edge and centre groups in each run and retains their labels. Which plan directly covers the named groups, and what still needs justification?

   Show the answer and reasoning

   Plan B directly covers them. Selection within groups, number of independent pieces, measurement quality, population weights and the intended precision still need justification.

   Thirty readings do not turn one location/run into the missing groups. Conversely, merely including each group once does not prove adequate precision or coverage of rare conditions.

4. Run A makes 100 parts and run B makes 300. Their supplied true mean thicknesses are 1.00 and 1.20 mm. Find the mean for a randomly selected part from their combined 400 parts, and compare the equally weighted mean of the two run means.

   Show the answer and reasoning

   The part-weighted mean is 1.15 mm; the equally weighted run mean is 1.10 mm.

   (100 × 1.00 + 300 × 1.20)/400 = 1.15 mm. (1.00 + 1.20)/2 = 1.10 mm answers an equal-run question. The supplied run means are assumed known here; sample estimates would add uncertainty.

5. For a cube of side 4.00 mm with the same supplied inward layer depth δ = 0.010 mm, calculate the altered volume percentage. Does it establish the percentage strength loss?

   Show the answer and reasoning

   1 − (3.98/4.00)³ = 0.014925125, or about 1.49%. It does not establish a strength loss.

   The result is a geometrical fraction under the uniform-layer assumption. Mechanical response depends on what changed, where the stresses act and how failure occurs; a small surface region can be consequential.

6. A new matched-coupon observation at 20 h gives 45% retained strength. Compare this with the supplied 60% linear and 44% alternative projections. Does it prove the alternative mechanism?

   Show the answer and reasoning

   It challenges the unqualified linear extension and is closer to 44%, but it does not uniquely prove the proposed alternative mechanism.

   Interpret the new value with its uncertainty, replication and matched state. Other curves or mechanisms may agree at that point. Additional times and relevant structural/failure evidence can discriminate explanations.

What to carry forward

• Define the population, sampled unit, measured quantity and selection process.
• Repeated readings, different locations and different runs supply different evidence.
• A change in size can alter geometry and sampled material state; state the particular model before transferring a property.
• A fitted model supplies predictions, not extra observations. Keep the observed range and proposed extension visible.