Sampling in Mineral Processing - Mineral Processing Hub

Sampling is one of the least glamorous parts of mineral processing and one of the most consequential. A poorly designed sampling system can make a profitable circuit look like it’s underperforming — or mask a real problem until it costs you a concentrate shipment. This guide covers what sampling is, why it goes wrong, and how to do it properly, from the theory behind sample size to the equipment you’ll see on a plant floor.

What Is Sampling and Why Does It Matter?

Sampling is the process of collecting a small portion of material that accurately represents the whole. In mineral processing, that “whole” might be a thousand tonnes of ore moving along a conveyor, or a slurry stream carrying concentrate to a filter. You can’t analyse all of it, so you analyse a fraction and trust that fraction to tell you the truth about the rest.

The stakes are high. Canadian Critical Minerals Research points out that accurate sampling directly influences resource estimation, feasibility decisions, and regulatory compliance — in other words, whether a project gets built, and whether it gets paid correctly. Get sampling wrong and you’re making every downstream decision on bad information.

A useful way to frame the problem: a large copper concentrator might process 100,000 tonnes of ore per day, yet the composite sample sent to the lab for the daily feed assay weighs about 1 gram. That gram needs to represent 1 part in 100 billion of a heterogeneous material. As Wills’ Mineral Processing Technology puts it, this is no trivial task.

Despite this, plants are often designed and commissioned with inadequate sampling infrastructure. Sampling equipment competes for capital against mills, flotation cells, and thickeners — equipment that feels more obviously central to production. The cost of that trade-off shows up later, in mass balance errors, disputed concentrate grades, and process control running blind.

For context on how sampling fits into plant design more broadly, see our guide to mineral processing plant flowsheet design.

Where Sampling Error Comes From

Sampling error has four distinct sources, and they add together rather than cancelling out. Reducing one source doesn’t offset another. You have to address each one on its own terms.

Composition error

This comes from the fact that individual particles are not identical. Some are richer in the target mineral, some are poorer, and they are not evenly distributed through the stream. The further the ore is from liberation — where mineral grains are still locked inside gangue — the more particle-to-particle variation exists, and the larger your sample needs to be to average that variation out.

This is the basis of Pierre Gy’s Theory of Sampling, developed in France in the 1950s and still the foundation of modern sampling practice. Gy showed mathematically that composition error depends on particle size cubed. Cut particle size in half and the required sample mass drops by a factor of eight. This is why crushing or grinding a sample before splitting it down for assay makes such a large practical difference.

Distribution error

Even if individual particles were identical, material in a stream is not randomly distributed. It segregates — by size, density, and shape. Coarse particles settle faster in a slurry pipe. Fine particles cling to belt edges. A single grab sample at one point in time, or from one location in a stockpile, cannot represent all of that variation. The solution is multiple increments spread across time and space, collected in a way that covers the full cross-section of the stream.

Preparation error

Once a sample is collected, it still has to be reduced to a size suitable for analysis. Crushing, splitting, drying, and splitting again — each stage introduces its own error if done carelessly. Rotary table splitters are the recommended tool for this stage because their geometry gives every particle an equal chance of entering each sub-container. Riffle splitters are prone to bias if the outside slots both feed the same side, or if the feeder isn’t perfectly centred.

Analysis error

The final source is the assay or measurement itself — instrument precision, calibration drift, operator variability. This is typically the smallest contributor. Merks (1985) estimated analysis error at around 2% of total error for coal sampling, compared to 54% from composition and 35% from distribution. Investing heavily in analytical precision while neglecting sample collection is one of the most common and expensive mistakes in metallurgical accounting.

How Much Sample Do You Actually Need?

Gy’s Theory of Sampling gives engineers a way to calculate the minimum sample mass needed to achieve a target level of precision. The key insight is that the required mass depends much more on particle size than on the total volume of material being sampled.

A few practical points that follow from this:

Gy’s calculation gives you a minimum for composition error only. Distribution and preparation errors sit on top of that. Doubling the calculated minimum is a common and reasonable starting point.
The material’s liberation size matters enormously. Ore types with fine liberation sizes — some copper porphyries, for instance — are more forgiving. Gold ores, where value is concentrated in rare, coarse particles, are the worst case. The “nugget effect” means that even a large sample can swing wildly if it happens to include or miss a single coarse gold grain.
Sample size requirements are not fixed — they depend on the precision you need. Metallurgical accounting for a concentrate shipment demands tighter tolerances than a process control measurement used to adjust reagent dosing.

For practical guidance on designing a sampling program for critical minerals projects, Canadian Critical Minerals Research outlines a project-tailored approach that factors in geological variability, mineral type, and regulatory requirements from the start.

The Ideal Sampling Model

Every particle in the stream should have an equal chance of ending up in the sample. That’s the principle behind what sampling theorists call the equi-probable or ideal sampling model. It sounds obvious, but most common sampling practices violate it.

A grab sample from the surface of a slurry tank violates it because fine, low-density particles accumulate at the surface. A poppet valve inserted into the centre of a pipe violates it because the velocity profile in a pipe is not uniform — the centre flows faster than the walls. A riffle split done slowly and unevenly violates it because some size fractions fall preferentially into certain slots.

Designs that come closest to the ideal model share a few common features: they cut the full cross-section of the stream from one side to the other, they do so at consistent speed, and they don’t exclude any part of the flow. A cutter traversing the full discharge from a conveyor head pulley, at constant speed, at random intervals, is a good approximation of the ideal. A probe inserted partway into a flowing pipe is not.

The practical implication: for any stream that matters to your metallurgical account, you need a sampler designed to meet the ideal model. Shortcuts on those streams don’t save money — they shift cost to assay disputes, reconciliation failures, and smelter deductions.

Running a Sampling Survey

Periodic sampling surveys give you a snapshot of plant performance. They’re used to benchmark the circuit, calibrate process models, identify losses, and support process improvement decisions. Running one well takes more planning than most teams allow for.

How many increments?

A composite sample built from multiple increments is more representative than any single grab. Statistically, precision improves with the square root of the number of increments — so going from 1 to 4 cuts error in half, and going from 4 to 16 halves it again. After about 8 increments the improvement per additional cut becomes small. In practice, 5 to 8 increments per composite is the accepted range for most plant surveys.

How long to run?

The survey needs to run long enough for all material currently in the circuit to have passed through the sampling points. The standard target is three times the mean residence time of the largest process volume in the circuit. For a grinding circuit with a 30-minute residence time, that means at least 90 minutes of steady-state operation before the survey data is considered valid.

Following any process change — a feed rate adjustment, a reagent switch, a mill stoppage — you need to wait the same duration before sampling again. Data collected during a transient period describes neither the old nor the new condition.

Check for steady state

Before using survey data, confirm that the plant was actually running steadily. Pull the process historian and look at mass flows, pulp densities, and key on-stream analyzer readings over the survey window. Cao and Rhinehart (1995) describe a useful computational method for steady-state detection. If key variables were drifting, the survey needs to be repeated.

The 911Metallurgist overview of plant sampling techniques is worth reading alongside this — it covers practical flowsheet considerations for choosing sample points and the trade-offs between automatic and manual sampling at different plant scales.

Between 30% and 50% of sampling surveys get rejected and re-run. Budget for this in your project schedule. Re-running a survey costs far less than making process decisions on data from a bad one.

Sampling Equipment

Sampling equipment falls into two categories: probabilistic and non-probabilistic. The distinction matters because probabilistic samplers are the only ones suitable for streams used in metallurgical accounting.

Probabilistic samplers

Linear samplers are the workhorse of plant sampling. A cutter traverses the full width of a stream — at a conveyor discharge point or at the end of a pipe — cutting a proportional slice of everything flowing past. The cutter moves at constant speed, starts and stops outside the stream, and uses knife-edge blades that deflect no particles preferentially to either side.

Heath & Sherwood’s linear sampler range covers the full spectrum of applications, from light-duty pneumatic units on small slurry pipes to the heavy-duty 1335 Niagara model for high-flow large-diameter discharge points. Their product documentation makes a clear case for why in-line probes and pressure pipe samplers — which don’t cut the full stream — produce bias that shifts unpredictably with changes in flow rate, slurry density, and pipe layout.

Vezin samplers are rotary devices used for secondary and tertiary sampling of smaller streams. The cutter sweeps through the full cross-section of a falling stream, taking a fixed percentage — typically 1.5% to 5% — on each pass. They’re a common choice downstream of a linear primary sampler, where the primary cut needs to be reduced further before going to the lab.

Rotary table splitters are the correct tool for dividing a collected sample into sub-samples. Multiple close-spaced radial containers rotate under a stationary feeder at constant speed, giving each container an equal share. Riffle splitters can work but are prone to systematic bias if not set up carefully.

For turnkey systems covering the full sampling chain from primary sampler to sample preparation, Multotec designs and manufactures equipment aligned with ISO, ASTM, AMIRA, and CNAM007 standards, with advisory services to help specify the right system for your application. They’ve installed over 700 sampling systems across international mining operations.

Non-probabilistic samplers

Pressure pipe samplers, gravity slot samplers, and in-pulp grab devices don’t cut the entire stream. They’re cheaper and easier to install than linear samplers, which is why they’re common — but they introduce bias that varies with operating conditions. They’re acceptable on process control streams where approximate readings are sufficient, but they should not appear on concentrate, feed, or tailings streams used for accounting.

Manual grab sampling with a bucket or in-pulp device is the worst case. It’s non-probabilistic, operator-dependent, and difficult to replicate. The 911Metallurgist article on plant sampling techniques discusses where hand sampling is still used in smaller operations and the practical limits of its accuracy.

On-Line Analysis

On-line analyzers measure process streams continuously or at short intervals, without the hours of delay involved in collecting a sample, preparing it, and sending it to the lab. They’re a key part of modern process control.

On-stream XRF

X-ray fluorescence analysis of slurry streams has been in commercial use since the early 1960s. A radiation source excites characteristic X-ray emissions from the elements in the slurry; a detector measures the intensity of those emissions; a calibration converts intensity to grade. The Outotec Courier system, now in its SL series, can handle up to 24 sample streams from a single centralized unit, cycling through each stream and reporting grades in near real-time. Critical streams get measured more frequently; tailings streams, which carry lower grades, get longer integration times for better precision.

In-stream probe systems do the same job without routing slurry to a central analyzer. A probe mounted in or near the slurry stream uses a compact radioactive isotope source and detector. These are more practical on plants where piping layouts make centralized sampling difficult, and they’re competitive with centralized systems in accuracy when combined with a well-stirred analysis zone.

On-belt PGNAA

Prompt Gamma Neutron Activation Analysis (PGNAA) measures the elemental composition of bulk material on a conveyor belt without touching it. A neutron source beneath the belt generates gamma emissions characteristic of each element in the material; detectors above the belt capture those emissions. The GEOSCAN-M is the most widely deployed system of this type, used in iron ore, copper, zinc-lead, manganese, and phosphate operations. Results are reported every two to five minutes, weighted by belt scale tonnage, and corrected for moisture by a microwave monitor built into the same unit.

Density and mass flow

Nucleonic density gauges measure slurry density continuously using a gamma source mounted on the outside of the pipe. The gamma beam passing through the slurry attenuates in proportion to density. Paired with a flowmeter — magnetic, ultrasonic, or array-based — the gauge gives a continuous dry solids mass flow rate. This combination is the standard approach for weighing slurry streams in circuits where belt scales aren’t practical.

For more on particle size measurement, which connects directly to both sampling program design and process control, see our complete guide to particle size analysis.

Mass Balancing

Sampling generates numbers. Mass balancing turns those numbers into a coherent account of where mass and metal are going in the plant. The two activities are inseparable — a mass balance is only as good as the sampling behind it.

The starting point is mass conservation: what goes in must come out. For a simple unit with one feed and two products — say, a flotation cell producing concentrate and tailings — you have two equations: the solid mass balance and the metal balance. From those two equations and three assays (feed, concentrate, tailings), you can calculate the mass split and the metal recovery without ever measuring the concentrate or tailings flow directly.

This two-product formula is simple and useful for quick checks. Its limits become apparent when you’re dealing with complex circuits, multiple streams, and measurement errors that don’t cancel neatly. A key sensitivity: recovery calculations become unreliable when the element you’re tracking isn’t well-separated between concentrate and tailings. Two circuits can show the same nominal recovery of 85% while one of them carries a 95% confidence interval of ±2% and the other carries ±20%, depending entirely on how wide the assay split is between the products.

Modern metallurgical accounting uses least squares data reconciliation rather than the two-product formula. The reconciliation adjusts all measured values — assays and flow rates together — by the smallest amounts necessary to make the data satisfy mass conservation across the whole circuit simultaneously. Measurements with tighter error estimates get adjusted less; measurements from problematic sample points get adjusted more. The result is a balanced dataset that respects what you trust and flags what you don’t.

Good reconciliation software also tells you, before you run a survey, whether your sampling design gives you enough information to estimate the variables you care about. A circuit can be globally over-determined — more data than equations — while certain individual streams remain impossible to estimate without additional measurements. Discovering this after the survey is expensive.

Practical Rules to Get Right

Most sampling problems on operating plants come back to a short list of mechanical and procedural failures. These are the ones worth checking first.

Cut the whole stream. The cutter opening must span the full cross-section of the flow, from edge to edge with nothing excluded. A cutter that clips the outside edges of a conveyor discharge, or a probe that only samples the centre of a pipe, will be biased in ways that change with operating conditions.

Keep cutter speed constant. A cutter that slows down as it passes through the stream takes a proportionally larger sample from the slow section and a smaller one from the fast section. This applies to manual and automatic samplers alike. For material with a top size above 1 mm, cutter speed should not exceed 0.6 m/s.

Size the slot correctly. The cutter slot must be at least three times the diameter of the largest particle, with a minimum of 10 mm. A slot that’s too narrow excludes coarse particles, biasing the sample toward fines.

Use knife-edge blades. Rounded or adjustable cutter blades deflect particles to one side. Knife-edge blades that are fixed and non-adjustable are the standard for probabilistic samplers.

Keep the chain of custody probabilistic. A probabilistic primary sampler followed by a non-probabilistic splitter produces a non-probabilistic result. Every stage of the sample reduction chain needs to preserve the equi-probable principle.

Match the sampler to the stream’s importance. Multotec and Heath & Sherwood both offer advisory services for specifying the right equipment for each stream’s role. Using a pressure pipe sampler on a final concentrate stream because it’s cheaper to install is a false economy.

Use the right standard. ISO, ASTM, AMIRA, AS, and CNAM007 all publish commodity-specific sampling standards. For accounting streams — concentrate shipments, feed to a smelter — compliance with the applicable standard is what makes your assay defensible in a commercial dispute.

Document everything. Canadian Critical Minerals Research emphasises that sampling quality assurance depends on written standard operating procedures, calibration records, and deviation logs as much as it depends on equipment design. Good documentation is also what protects you when a counter-party challenges a shipment assay.