Process models are often fitted to measurements whose nominal precision is far better than their practical reproducibility. An assay instrument may report a grade to four figures, while successive survey composites differ by a tenth of a percentage point because coarse mineralized fragments occur sporadically. The relevant uncertainty includes the extracted material, sample preparation and analytical determination. An equipment model can respond to those errors in the same direction as it responds to a true change in ore. A successful plant study has enough information to tell those explanations apart.
Define what a composite estimates
An increment is selected from a specified material population. A set of increments can represent a campaign, a stockpile lot or a cross-section during a process survey. A representative composite is an estimator of a declared quantity. Time-uniform sampling with equal increment masses estimates a time mean if each increment captures the same effective cross-section. For total metal divided by total dry mass over a campaign, a flow-proportional composite or sufficiently resolved individual rates and grades is required. The distinction grows when the high-throughput period has a different ore grade.
Cross-stream extraction seeks to give particles across the moving stream an opportunity to enter the sample. Segregation across a belt or launder becomes a bias if the extraction geometry favours the near edge or the surface. Two duplicate samples can reproduce that bias closely. Repeatability under a flawed extraction method is limited evidence about the intended population. A review can compare a survey stream against an independent whole-stream measurement, planned belt-cut checks or suitable higher-quality sampling at a location where the same material can be identified.
Pierre Gy's sampling framework separates error arising from particulate heterogeneity from errors introduced by how material is selected and handled (Gy 1976). This separation has practical meaning. A larger mass moderates random variability when a few coarse grains carry a large share of the tracer. It cannot correct systematic exclusion of an unobserved portion of the cross-section. A well-selected increment can still be biased by retained wet fines, contamination, attrition or altered sample moisture during preparation.
Independent replication and a simple variance budget
Suppose a plant assay result is affected by variation between primary increments, sample preparation and instrument repeatability. For a balanced illustrative design, assume the three contributions are independent, mean zero, and have variances σ²ₚ, σ²ᵣ and σ²ₐ on the same grade units. If each primary increment is prepared separately, with r preparations per increment and a independent final determinations per preparation, the grand mean based on n increments has the approximate variance below.
For a tracer measured in percentage points, let the estimated variance components be 0.12², 0.06² and 0.02². Four increments, two preparations and two readings per preparation give 0.003600 + 0.000450 + 0.000025 = 0.004075 percentage points squared. The standard error is about 0.0638 percentage points. Doubling laboratory readings halves only the last term and reduces the standard error to about 0.0637. Doubling independent primary increments would instead produce approximately 0.0451. This calculation is a study-design comparison under independent sampling. It does not generalize to systematic effects or strongly correlated increments without adjustment.
For autocorrelated observations, the variance of the mean contains covariance contributions. Positive serial dependence is common as a stockpile parcel or feed interval takes time to pass. In a simple equally spaced stationary sequence, separations between increments need to be sufficiently large or the effective sample size is lower than the number collected. Densely logged process data can reveal useful timing and variation patterns while providing comparatively little independent evidence about the long-run mean or a conditional model fit.
Closure and uncertainty in a sieve run
Treat the total dry test mass before sieving and the recovered masses after sieving as separately meaningful records. The ratio of the sum of recovered dry masses to the initial dry mass gives test recovery. Reporting percentages on recovered mass makes a convenient normalized vector. It also causes a shared denominator and shifts all values if one class suffers disproportionate loss. A loss concentrated on the finest sieve and pan can make a nominally clean fractional distribution materially wrong.
At a 250 µm screen aperture, a test measures 6.50 kg above and 2.90 kg below from 9.60 kg of initial dry mass. Recovery is 97.92%. The fraction passing on recovered mass is 2.90/9.40 = 0.30851. A deliberately conservative bound places all unaccounted 0.20 kg on either side. The resulting range on initial mass is [2.90/9.60, 3.10/9.60] or [0.30208, 0.32292]. If elevated dust is observed during transfer, the fine-side endpoint is better motivated than an assumption of proportional loss. When the downstream screening limit is 31.0%, the current test cannot decisively distinguish the acceptable state.
Coarse size assays may use relatively large sample masses because a few large grains carry appreciable weight. A residue from mineralogical mounting or drying is a different sample with its own selection process. If the training data combine assays from different subfragments of a sample, preserve the shared original identifier and distinct preparation notes. Calibrating a model from a merged table with opaque provenance makes later disagreements hard to diagnose.
Detection limits and zero mass
A grade reported below a detection limit represents a censored value. A blank cell may represent an uncollected assay. A zero dry flow can make grade undefined. The numbers that reach a database often look alike, but the engineering interpretation differs. If a feed contains 150 t/h dry solids and a tailings rate of 120 t/h with a grade reported below 0.02%, the observation implies less than 0.024 t/h of tracer in tailings under that threshold. It does not identify a zero loss. Censoring can be represented as a bound or a censored sampling likelihood, depending on the information available.
When normalizing grades on a complete mineral basis, define how residual gangue is constructed. The residual has correlated error with every quantified constituent. For a small tracer grade on a background of several classes, the estimate may be nearly independent of unmeasured material, while an inferred remainder grade depends on all preceding measurements. Reconciliation with independent bulk and within-size analyses must preserve these dependencies if uncertainty estimates are to have physical meaning.
Sampling support and a rejectable prediction
Imagine comparing two cyclone models under equal steady feed. Both give a corrected cut size of 100 µm, while they predict 11% and 16% fine underflow bypass. A duplicate survey using repeated studentized differences estimates bypass of 14.5 ± 1.5 percentage points as an approximate one-standard-error interval. The 11% prediction differs by 2.33 standard errors and 16% by one. If those uncertainty assumptions hold, this measurement is relatively more consistent with the second model. If both readings came from poorly timed feed and underflow samples during start-up, the survey has an additional timing uncertainty that makes this simple comparison optimistic.
The finest size fractions are often most useful for distinguishing bypass from corrected classification. They are also prone to low dry masses and measurement loss. A targeted survey should improve that specific part of the evidence. A bulk throughput campaign with twice as many time points may do little to distinguish the two fitted fine-end structures.
Exercises with solutions
Exercise 1. A two-hour survey observes 100 t/h at a 0.8% tracer grade in the first hour and 160 t/h at 0.5% in the second. Calculate the grade of the full material lot and the mean of two equal-mass time-composite increments.
Both hours contribute 0.8 t of tracer. The full material lot contains 1.6 t tracer and 260 t solids, giving 0.61538%. The equal-mass time-composite estimates 0.65%. Which is the intended observation must be declared before comparison with a campaign accounting model.
Exercise 2. A flotation concentrate grade has independent estimates of 23.75%, 24.10%, 23.95% and 24.20%. Estimate the unweighted mean and sample standard error under an independent, equal-variance assumption.
Mean grade is 24.00%. Squared deviations sum to 0.115 percentage points squared. Sample standard deviation is √(0.115/3) ≈ 0.196. The estimated standard error is 0.098. The four values support a statement about replicate analytical precision on the tested material; sampling variance may require independent primary composites.
Exercise 3. The component rate is calculated from a 160 t/h feed with a 1.2% tracer grade. Independent standard uncertainties are 5 t/h and 0.10 percentage points. Estimate first-order uncertainty in tracer flow.
The calculated component flow is 1.92 t/h. The sensitivity to total flow is 0.012 and to grade on the fraction basis is 160. The approximate variance is (0.012 × 5)² + (160 × 0.001)² = 0.0292 (t/h)². The resulting standard uncertainty is about 0.171 t/h. Correlation between estimated grade and rate would add a covariance term.
