eleticle

Core No. 01Logged 6 Oct 2026

Recovery moved by two percentage points. Did the circuit actually improve?

Worked example · constructed grades · no site data

A worked example with made-up assays for one copper flotation circuit. We keep the plant exactly as it is, change only what the lab reports, and recalculate recovery with the standard two-product formula.

What we found

  • An assay shift looks like a gainWith nothing in the plant changed, a tailings assay reported at 0.08 instead of 0.10 % Cu lifts calculated recovery by 1.92 percentage points.
  • A real drop looks the sameIf the tailings grade really fell to 0.08 %, recovery would read the same 92.37 %. The formula cannot tell a better circuit from a shifted measurement.
  • A small assay, a large shift0.10 to 0.08 % Cu is 0.02 percentage points of grade, and 20 % of the assay itself.
  • Not only tailingsRaising the feed or the tailings assay by 5 % moves recovery about half a point, up for feed and down for tailings. The concentrate assay moves it least.
  1. Method
  2. 01The report
  3. 02The calculation
  4. 03Three reports
  5. 04The other assays
  6. 05Making it credible
  7. 06Before crediting

Method

A constructed circuit and a standard formula

  • ConstructedEvery grade, flow and recoveryIllustrative values chosen to make the mechanism visible, not data from any site. They show how the formula responds, not how often or how far real assays err.
  • EstablishedThe two-product formulaRecovery from three assays, feed, concentrate and tailings, through the solids and copper balances. The example is a worked application of it, not a new result.
  • CheckedExact arithmeticThree grades and the formula give every value, by hand or in a spreadsheet. Our short script recomputes them and the balances at 100 dry t/h; the curve and the slider use the same function.

The references are under Sources, at the end of the page.

01 · The reported improvement

Recovery is up. What does the number establish?

A copper flotation circuit changes its collector dosage. The next month's report shows recovery at 92.4 %, up from 90.5 %. That is the kind of movement that gets a trial written up as a success. Before it is, three separate questions need an answer, and the recovery number on its own answers none of them.

  1. Is the calculation right?Which definition, which boundary, which inputs.
  2. Is it bigger than the measurement can resolve?Uncertainty in both periods, and in the difference.
  3. Did the change cause it?Ore, throughput and everything else that moved.

02 · The calculation

What the recovery number is built from

Draw a boundary around the circuit with one feed, one concentrate and one final tailings stream. Recycles between cells may exist, as long as they stay inside the boundary and every stream that crosses it is counted. Treat the boundary as steady state, with no change in stock.

Feed, F and f, enters a dashed boundary around the flotation cells. Concentrate, C and c, and tailings, T and t, leave it. A recycle between the cells stays inside the boundary.
F, C and T are dry-solids mass rates. f, c and t are copper grades of the same streams, on the same dry basis, from samples over the same period.

Solids and copper are both conserved across the boundary:

F=C+TFf=Cc+Tt

Put T = F − C into the copper balance and divide by F for the concentrate yield:

Y=CF=f−tc−t

Recovery is the share of feed copper that reports to concentrate:

R=CcFf=c(f−t)f(c−t)

Fractions, one scale

Y and R are fractions; multiply R by 100 for percent recovery. The grades can all be fractions or all percentages, because their common scale cancels. Mixing the two is an error.

An estimate from three assays

This is an assay-derived two-product estimate, not a reconciled plant balance built on measured flows and several components. For a normal enriching separation, as here, c > f > t ≥ 0. When the inputs break that, or recovery lands outside 0 to 100 %, look at the inputs and the boundary. Do not clip the result.

These grades are the example's true condition. With 100 dry t/h of feed they imply 4.52 t/h of concentrate and 95.48 t/h of tailings, which close both the solids and the copper balance. Recovery is 90.45 %.

03 · Same plant, three reports

Change only the reported tailings assay

Hold the plant exactly as it is. Keep the reported feed and concentrate at 1.00 and 20.00 % Cu. Change only the reported tailings assay, as a measurement might shift. Nothing in the circuit has moved, and none of this is an operating trial.

Calculated recovery · constructed exampleReported 0.10 % Cu · recovery 90.45 %
Calculated recovery against the reported tailings grade from 0.06 to 0.14 % Cu, feed 1.00 % and concentrate 20.00 % held. The line falls almost straight from about 94.3 % to 86.6 %. A reported 0.08 % gives 92.37 %, 1.92 percentage points above the reference of 90.45 % at the true 0.10 %; 0.12 % gives 88.53 %, 1.92 points below.

A deterministic sensitivity curve for a constructed example. Every point is the same true plant, with feed and concentrate assays held. It is not a confidence band and not a time series.

Reported tailingsRecoveryChangeImplied yield
0.08 % Cu92.37 %+1.92 pts4.62 %
0.10 % Cu90.45 %0.00 pts4.52 %
0.12 % Cu88.53 %−1.92 pts4.43 %

A reported 0.08 % Cu gives 92.37 %: the “after” figure in the report above. From a tailings stream that has not changed, the formula has inferred more concentrate, 4.62 % of the feed instead of 4.52 %, and so more copper recovered.

0.02 points is not a small error

From 0.10 to 0.08 % Cu is 0.02 percentage points of grade, but 20 % of the assay itself. At low tailings grades, a small absolute shift is a large relative one. The range was chosen to make the effect visible. It is not a claim about how well any laboratory or sampling system performs.

Points, not percent

Recovery went from 90.45 to 92.37 %. That is 1.92 percentage points, or a 2.1 % relative increase. A report should say which it means.

A real drop looks the same

If the true tailings grade had fallen to 0.08 %, the calculation would give the same 92.37 %, and it would be right. The formula cannot tell a better circuit from a shifted measurement. That is the problem, not a verdict on any reported gain.

04 · The other assays

Check the other assays too

Tailings is not always the assay that matters. Start again from 1.00, 20.00 and 0.10 % Cu, move one assay at a time by 5 % of its own value, and hold the other two.

Recovery, %, at −5 % and +5 % of each assayHollow −5 % · filled +5 %
Reference recovery 90.45 %. Feed assay 0.95 % gives 89.92 %, 1.05 % gives 90.93 %. Concentrate assay 19.00 % gives 90.48 %, 21.00 % gives 90.43 %. Tailings assay 0.095 % gives 90.93 %, 0.105 % gives 89.97 %.

Each label is the assay input, % Cu, and the recovery it gives. These are deterministic perturbations, not standard deviations, confidence bounds or uncertainty estimates.

Assay moved−5 % input
→ recovery
+5 % input
→ recovery
Feed, 1.00 %0.95 %
→ 89.92 %
1.05 %
→ 90.93 %
Concentrate, 20.00 %19.00 %
→ 90.48 %
21.00 %
→ 90.43 %
Tailings, 0.10 %0.095 %
→ 90.93 %
0.105 %
→ 89.97 %

At this operating point, an equal relative change in the feed or the tailings assay moves recovery by about half a point either way. The concentrate assay moves it by less than a tenth of a point. How much each assay adds to real uncertainty depends on how uncertain it actually is and how its errors move with the others, so this ordering does not carry over to every separator or ore.

What the literature finds

Iveson and Galvin (2024) derive how the two-product formula amplifies assay uncertainty. Under their assumptions they also find the concentrate grade has little influence, and the gap between the feed and tailings grades governs.

Under stated assumptions

Their simple forms assume the same relative uncertainty on all three assays, independent random errors and small errors, and do not apply to mass-balanced data. They show where to look, not a rule for every circuit.

05 · Making it credible

What would make the improvement credible

The example shows that a measurement effect can imitate an improvement. It does not show that any particular reported gain is false, insignificant or caused by assay error. That takes evidence of three kinds.

Sensitivity is not uncertainty

The curve shows what the formula does when an input changes. Measurement uncertainty needs evidence about the whole chain: primary sampling, preparation, analysis, timing and any systematic effects. Repeat assays of the same prepared pulp describe the laboratory step only. Gaylard and co-authors (2014) set metal accounting on three legs, mass measurement, sampling and preparation, and analysis, each of which must be free of bias.

Judge the difference

The uncertainty of a before and after change depends on the uncertainty of both estimates and on how they are related. Add the two variances only when the two estimates are uncorrelated, as they are when their errors are independent. A bias common to both periods behaves differently in the difference from fresh random error in each, and samples taken close in time are often correlated.

Detection is not attribution

A change that is well resolved still need not come from the intervention. Feed mineralogy, liberation, throughput, residence time, sampling windows and other operating changes can all differ between the two periods. Designing a trial that separates them is a topic of its own.

Method box · the uncertainty of a change

ΔR=Rafter−Rbefore u2(ΔR)=u2(Rafter)+u2(Rbefore)−2Cov(Rafter,Rbefore)

Here u is a standard uncertainty and Cov the covariance of the two recovery estimates, all in the same recovery units. This is the first-order law of propagation, applied to a difference (NIST TN 1297, Appendix A). Drop the covariance term only when a zero covariance is justified; independent errors are one way to justify it. An interval quoted for ΔR should state the input uncertainties assumed, whether they are standard or expanded, the distribution and covariance assumed, and how coverage was found. The sweep above is none of these, and it gives no two-point threshold.

06 · Before crediting the change

Six checks on a reported recovery gain

  1. Which recovery, from which calculation?The definition, the formula and its inputs: a two-product estimate or a reconciled balance.
  2. Is the boundary closed?Every external stream, any change in stock, and whether steady state holds over the period.
  3. Do the samples match?Comparable material and time windows. Go by when a sample was collected, not when the laboratory reported it.
  4. What does the whole measurement chain show?Primary sampling, preparation, analysis, known bias and quality-control evidence. Repeating an assay does not repair an unrepresentative primary sample.
  5. How uncertain is the difference?The uncertainty of the before and after change, with dependence assumptions you can justify and repeats that represent the whole chain.
  6. What else changed?Ore, throughput, residence time and other operating changes, before the gain is credited to the intervention.

A two-point gain in recovery can matter. The number alone does not establish whether it is real or what caused it. Give it a measurement check and a process-context check before crediting it to the intervention.

Sources

  1. Iveson, S. M. and Galvin, K. P. (2024). Amplification in the uncertainty of the yield and recovery for a steady-state mineral separator calculated using the two-product formula. Minerals Engineering, 212, 108729. doi:10.1016/j.mineng.2024.108729. Open access; the publisher version was read.
  2. Virginia Tech, MinE 3544 Mineral Processing Laboratory. Lab 01: Laboratory calculations, teaching notes, updated 15 January 2020. PDF, used to check the balance, yield and recovery definitions.
  3. Gaylard, P. G., Randolph, N. G. and Wortley, C. M. G. (2014). Metal accounting and corporate governance. Journal of the Southern African Institute of Mining and Metallurgy, 114(1), 83–90. Full text.
  4. Taylor, B. N. and Kuyatt, C. E. (1994). Guidelines for evaluating and expressing the uncertainty of NIST measurement results. NIST Technical Note 1297, Appendix A, equation A-3. PDF.