Taking a Steel Sample for Laboratory Testing
When a customer disputes a composition, an inclusion count or a case depth, the argument is usually about the sample rather than about the laboratory. Sampling has its own uncertainty, separate from the uncertainty of the measurement, and on a material as variable as a steel bar it is commonly the larger of the two. This page covers what a sampling plan has to decide, how the spread in a reported figure divides between the sample and the method, and the specific sampling rules that apply to metal, including the ones that are easy to break without noticing.
The sample is a measurement in its own right
Sampling is the selection of a portion of a population for testing, and the reliability of any result depends on the quality of that portion. A poorly designed plan, uncertainties in the taking, or a problem in storage, preservation or pretreatment can obscure a result or make it impossible to interpret, and none of the controls a laboratory applies afterwards will recover it. Blanks, standards and reference materials control the other error sources, and they cannot fix an invalid sample.
The practical consequence is that sampling uncertainty is treated separately from the rest. The overall standard deviation of a reported figure is the sum of the sampling contribution and the analytical contribution, and on a heterogeneous material the sampling term can dominate. That is why a well run laboratory can return a precise result that is still wrong about the parent lot, and why the sample has to be argued about before the number is.
Where the spread actually sits
A result taken from a lot of material has three contributions to its spread, and they respond to different decisions.
| Where the spread sits | What it is | How it is reduced |
|---|---|---|
| Between units in the lot | The genuine variation from bar to bar, bundle to bundle or car to car | Sampling more units, which is the term the worked example below reduces first |
| Between samples inside one unit | Variation within a single bar, coil or bundle, which is where segregation inside a heat shows up | Taking more samples from each unit that is chosen, and taking them systematically rather than where they are easy |
| The analytical operation | The repeatability of the laboratory method itself, including the repeatability of any subsampling it does | Replicate analyses, which is the most expensive route per unit of variance removed |
The variance of a reported composition figure splits into these parts. Compiled from the source article in ASM Handbook, Volume 10.
The rule that follows is the one worth remembering from the whole subject. Once the measurement uncertainty has been brought down to a third, or less, of the sampling uncertainty, further refinement of the measurement is unimportant. Where the sampling uncertainty is large and cannot be reduced by better sampling, a rapid approximate method may be entirely sufficient, and running that method over more samples will reduce the uncertainty in the average more than upgrading the instrument would.
What the plan has to decide, and why the words matter
A sampling plan fixes the number, size and location of the increments, whether they are analysed individually or blended into a composite, and how the gross sample is reduced to a laboratory sample. The distinctions between the kinds of sample are not hair splitting, because each of them supports a different claim.
A random sample means every part of the population has an equal chance of being taken, and it is collected from a table of random numbers rather than haphazardly. Any sample selected by a defined protocol reflects the biases of the sampler, so the apparently unsystematic pattern has to be followed exactly for the result to hold. A representative sample is a different idea. It means one sample expected to show the average properties of the population, and the article is clear that such a sample cannot be selected by a random process and cannot be verified as representative afterwards. A composite sample is a special case of it, produced by crushing, grinding and blending, and it has a real cost. Analysing individual samples gives the average and the variation between samples, while a composite gives only the average, and the choice between the two should be made deliberately rather than for convenience.
The number of increments is not a judgement call either. A small preliminary set is analysed, the standard deviation of the individual samples is calculated from it, the confidence limits on the average are established, and a refined plan follows from those figures. One or two cycles of that usually pins the parameters down well enough to design the sampling programme with known confidence.
| Relation | What it tells you |
|---|---|
| overall standard deviation squared = sampling standard deviation squared + analytical standard deviation squared | The two contributions to the spread in a result are separable, and they should be measured separately wherever that is possible |
| reduction in measurement uncertainty stops paying once it is one third or less of the sampling uncertainty | On a material that is genuinely variable, a fast approximate method plus more samples beats a slower, better method |
| w R2 = Ks | The sampling constant. It is the weight of a single sample needed to hold the sampling uncertainty to 1 % at 68 % confidence, and it only applies to a well mixed sample with no segregation |
| n = t2 σs2 / E2 | The number of increments needed. t is the Student value for the confidence wanted, 1.96 at 95 %, and E is the standard deviation acceptable in the average |
| σs2 = A / w n + B / n | The form for a segregated material. A is a homogeneity constant and B a segregation constant, both determined experimentally for the material in question |
The working relations behind a sampling plan, in the order they are used. Source, Sampling, in ASM Handbook, Volume 10.
A worked example on metal
The source material includes a worked example on a trainload of metal pipe being sampled for the percentage of an alloying element, and it is instructive because it shows where the money goes. The standard deviation was 0.25 between cars, 0.15 within a car, and 0.08 for a determination, with relative costs of 5, 4 and 1 for sampling a car, taking a sample within a car, and running an analysis.
| Goal | Optimum scheme | Resulting spread | Relative cost |
|---|---|---|---|
| Hold the overall standard deviation at or below 0.10 | 11 cars sampled, 1 sample per car, 1 determination per sample | 0.09 | 60 |
| Spend no more than 40 on the relative cost scale | 5 cars sampled, 1 sample per car, 1 determination per sample | 0.14 | 50 |
A trainload of metal pipe sampled for the percentage of an alloying element. The spread between cars is 0.25, within a car 0.15, and between determinations 0.08, with relative costs of 5 to 4 to 1. Source, Example 1 of Sampling, in ASM Handbook, Volume 10.
Both rows in that table support the same conclusion. When the spread between units is the largest term, spending the budget on more units is what buys accuracy, and a single determination per sample is already enough. Adding replicates inside one car cannot compensate for having sampled too few cars, and this is the mistake behind a very large share of disputed test results.
Rules that apply specifically to steel
Metal looks uniform in a way that ore and powder do not, and that appearance is misleading. The article notes that even where the material is uniform in structure, chips or particles of different sizes produced by crushing or machining can have significantly different compositions. Drillings, millings and saw cuttings from the same bar are not interchangeable samples, and one fine fraction of the same cut is not interchangeable with the coarse fraction.
Cutting is the second trap. Using a cutting method to remove a sample from a metal structure for metallographic examination is to be avoided, because the heat generated by cutting can alter the microstructure significantly, and a torch cut is the extreme case. Where a torch has to be used, a large sample is taken and the metallography is done at a location the cut has not reached. More generally, a sample has to be taken in a way that does not influence the measurement being made, and the article records that many incorrect conclusions have been drawn because the sample was altered while it was being taken.
Sampling devices carry the same logic. A sampling tool should not be made of, or contain, the element being determined, which is why a plastic scoop is the right tool for a metal sample and a metal scoop is the right tool for an organic one. Where equipment has critical dimensions or settings, those have to conform to the protocol, and tools that can drift, such as pumps and sieves, need their calibration checked at the time of use.
Handling, identity and subsampling
Composition can move between taking and testing, through oxidation, differential evaporation, loss of volatile constituents, thermally induced change, or interaction with the container. Airborne dust is a real source of contamination in trace work. Preservation options include low temperature storage, an inert atmosphere, hermetic sealing and opaque containers, and the test of whether a preservation method works is a holding time trial, in which preserved samples are analysed periodically until an intolerable change appears.
Two further points are worth carrying into any discussion with a laboratory. Subsampling is sampling the sample, and unless there is proof to the contrary its error should be assumed to be present, so a laboratory that takes a portion out of the sample it was sent has added a step with its own uncertainty. And ordinary cleanliness matters more than it sounds, because equipment reused between samples can carry contamination or memory from the last one, and cleaning procedures may themselves be hard to verify. Sample identity is the last item on the list, and the requirement is that there should be no reasonable doubt about the identity and integrity of anything analysed.
What this means for a tool steel order
Three practical habits follow. Take the sample from the location that carries the risk, which for a rolled or forged bar means thinking about where segregation and where the surface condition are worst, and record where it came from rather than just labelling it. Ask the laboratory how much it needs, in what form and how many pieces, before cutting anything, because a wrong sample form costs a repeat test and a delay. And when a figure is disputed between two laboratories, separate the two questions before assuming either is wrong. The method’s own repeatability can be checked with a replicate, and the difference between two samples from the same lot can be checked with a second sample, and those two tests will point at the source.
Sampling also sits underneath the two measurements most often argued about on tool steel. The rules on polished surface area and field count for an inclusion or grain size result are set out in measuring inclusions and grain size in steel, and the methods behind a chemistry certificate are covered in how the chemistry of a tool steel heat is tested. The defects a sample can reveal in the bar itself are collected in steel bar defects, our own incoming and outgoing checks are described in how we control tool steel quality, and when a sample is being cut for a failure investigation the depth and location advice sits on that same metallography page.
Grade families with their own delivery conditions are gathered under cold work and hot work tool steels, the catalogue is at tool steels, a grade can be matched with the tool steel material finder, and a sampling or testing question can be put to us through the contact page.
Before you cut a sample
A reference page, it is not an Aobo Steel specification, and the figures above are the published values and relations for the case described. A test result belongs with the sample it came from, so record the location, the form and the number of pieces, and keep the cut cool where hydrogen or oxygen is being measured. Where a result is disputed, check the sampling uncertainty before treating the difference as a disagreement about the steel.
Source: ASM Handbook, Volume 10, Materials Characterization, ASM International, 1986.