For a mineral engineer tackling a new problem (e.g., "Why is recovery dropping in the rougher cells?"), follow this statistical workflow:
Pierre Gy famously stated: “No amount of statistical processing can correct a bad sample.”
The Fundamental Sampling Error (FSE): [ \sigma^2_FSE \propto \left( \frac1M_S - \frac1M_L \right) \cdot f \cdot g \cdot c \cdot d^3 ] Where:
Practical implication for mineral engineers:
Shocking fact: Over 50% of plant metallurgical balance errors originate from poor sampling, not poor analysis.
Professor Amaya Calder had taught statistical methods for mineral engineers long enough to know the stubborn rhythms of rock: how randomness and pattern braided through the earth like veins of ore. Her classroom smelled faintly of coffee and chalk and, on stormy afternoons, of wet soil tracked in by students who’d come straight from the pit.
When the mining company announced the new high-grade deposit at Cerro Viento, the regional team called her in. The deposit’s assay data were messy: clusters of high values, long tails of low-grade samples, and pockets where grade rose and fell with little warning. Investors wanted a single confident estimate of recoverable metal. The foreman wanted a drill plan. Politicians wanted reassurance that the mine wouldn’t poison the groundwater. And Amaya wanted to teach her students one more lesson — that sound decisions begin where curiosity collides with uncertainty.
She arrived at the site with a battered field notebook and a laptop full of scripts. Her graduate assistant, Lin, a meticulous thinker who could coax patterns out of chaos, met her by the core shack. They unfolded sample logs into a patchwork of numbers: sixty cores, each cut into half-meter intervals, each interval carrying an assay. The raw histogram looked like the craggy skyline of a mountain range — peaks, troughs, and long, ragged tails.
“People will want averages,” Lin said. “But the mean will be dragged by those outliers. If we present that, we’re lying by decimal point.”
Amaya smiled. “Statistics isn’t a single number handed down from on high. It’s a conversation. We choose methods that match the rock and the questions.”
Their first step was exploratory data analysis. They plotted boxplots and rank-order graphs, looked for skew, and mapped the spatial coordinates of samples. The high-grade clusters weren’t uniformly distributed; they traced a loose lens dipping to the east. Some assays flagged as extreme, but when mapped they fell into a continuous filament—likely real structure, not lab error.
They tested for normality and quickly rejected it. The grade distribution was log-normal with heavy tails. Amaya suggested a log-transform for many analyses but warned against blind application. “Transformations help with modeling, not with telling the whole story,” she said. “We have to interpret back in original units for engineering decisions.” Statistical Methods For Mineral Engineers
Next came variography: semivariograms, nugget effects, and range. These tools measured how similarity decayed with distance. Lin calculated experimental variograms in multiple directions. The anisotropy was clear: correlation extended farther along strike than down-dip. That mattered for kriging—an interpolator that weights nearby samples according to spatial correlation.
They built nested variogram models: a small nugget to capture sampling and microscale variability, a short-range spherical structure for pocket-scale continuity, and a longer-range exponential structure for broad-grade trends. With the models fitted, ordinary kriging produced smoothed grade estimates across the block model, but Amaya knew kriging’s smoothing bias could underestimate high-grade variability — dangerous for resource classification and project economics.
“Use conditional simulation,” she told Lin. “We need realizations that honor both the data and the variogram, so we can quantify uncertainty for each block.”
They ran sequential Gaussian simulation, generating dozens of equally probable 3D realizations of the grade field. Each realization preserved the global distribution and spatial continuity while allowing high-grade clusters to appear or vanish in different places. Together the realizations painted a probabilistic landscape: the probability that a block exceeded economic cutoff, the range of possible recoverable tonnages, and the worst-case scenarios investors dreaded.
With simulations in hand, they computed conditional cumulative distribution functions for key pitshells. Decisions stopped being yes-or-no and became questions of acceptable risk. The mine planner could choose a conservative cut-off to ensure high confidence in early cash flow, or a riskier approach that chased upside while hedging with phased development.
Amaya also insisted they look beyond grade. Bulk density varied with lithology. Recovery rates depended on mineral liberation characteristics the assays didn’t capture. She introduced multivariate techniques: principal component analysis to summarize correlated geochemical indicators and co-kriging to incorporate secondary variables where appropriate. For zones with scarce sample density, they used indicator kriging to estimate the probability of crossing critical thresholds rather than trying to estimate a precise mean.
The students watched as statistics moved from abstraction to consequence. One night, a younger engineer named Mateo asked, “Which method is right? Kriging, simulation, indicator—how do we pick?”
Amaya wrote a short list on the whiteboard:
“Every method has limits,” she said. “But when we combine them judiciously, they form a fuller picture.”
Weeks later, the company faced a decision: expand the pit now, risking early capital for uncertain high-grade pockets, or stage expansion after additional drilling. The board asked for a recommendation. Amaya prepared a concise report: maps showing kriged grade means, probability maps from simulations, sensitivity analysis of recoverable metal under different cut-offs, and the economics under several scenarios. She highlighted blocks with high probability of exceeding cutoff but high conditional variance — the places where an extra drill hole would most reduce uncertainty.
Her recommendation was both statistical and pragmatic: proceed with a phased expansion focused first on blocks with high mean and low uncertainty; defer high-variance, high-upside blocks pending targeted infill drilling. Include a monitoring program to update models as new data arrived. Tie early production decisions to probabilistic thresholds rather than fixed arbitrary numbers. For a mineral engineer tackling a new problem (e
The board approved the phased plan. Investors liked the transparency. The foreman liked the clear priorities for drilling. And the environmental officer appreciated that uncertainty quantification reduced the risk of surprises that could endanger water or nearby communities.
A year later, after a season of follow-up drilling, the updated simulations tightened. Some high-variance blocks resolved as true bonanzas; one promising filament proved barren. The phased strategy’s flexibility—rooted in sound statistical thinking—saved millions in sunk capital and avoided disruptive mid-project pivots.
On the last day before she returned to teaching, Amaya walked the site with Lin and Mateo. They stood on a low ridge and looked across the grid of boreholes, the checkerboard of samples, the pit outline traced by engineers and statistics alike.
“You taught us the math and the models,” Lin said. “But more than that — you taught us to treat uncertainty like information, not an obstacle.”
Amaya watched the clouds move slow and indifferent over the mountain. “Rocks don’t care about our plans,” she said. “They simply are. Statistics lets us listen.”
Back at the university, her next semester’s syllabus changed slightly. She added a practical module: students would build kriging models, run conditional simulations, and present risk-informed mine plans. She sent her class into the world with notebooks and scripts, but also with a quiet creed: measure carefully, question boldly, and always make decisions that respect both data and uncertainty.
In the years that followed, some of her students led projects across the globe. Each time they faced a stubborn deposit, they remembered Cerro Viento — not as a triumph over nature but as a lesson in partnership with it. The ore remained patient and variable; the engineers became better at asking the right questions, and the decisions made from their statistics were, more often than not, wiser.
End.
Statistical Methods for Mineral Engineers Mineral engineering is increasingly defined by the complexity of lower-grade ore bodies and the demand for higher operational efficiency. In this environment, statistical methods serve as essential tools for transforming raw plant data into actionable intelligence, allowing engineers to optimize recovery, manage uncertainty, and make data-driven decisions. 1. Fundamentals of Data Analysis in Mineral Processing
At its core, statistical analysis for mineral engineers begins with understanding the variability inherent in geological and processing data. minerals - SBUF
You can use this as a LinkedIn article, a blog post, or a technical memo. Practical implication for mineral engineers:
Linear regression is the workhorse, but mineral processes are rarely linear.
Pierre Gy dedicated his life to the statistics of sampling. His fundamental law is that the sampling variance (apart from geological variance) is inversely proportional to the sample mass.
Gy’s Formula for Fundamental Sampling Error:
$$ \sigma^2_FSE = \frac1M_S \left( \fracf g \beta d^3c \right) $$
Where:
The Golden Rule for Mineral Engineers: For a given desired variance, if you double the particle size ($d$), you must increase the sample mass by 8 times ($2^3$).
Practical Application: You are designing a sampling protocol for a leach feed. The grind size is $P_80 = 75 \mu m$. You take a 200g pulp for analysis. The variance is acceptable. Now you need to sample crushed ore at $P_80 = 10mm$ (10,000 $\mu m$). The particle size ratio is $10,000 / 75 = 133$. The mass required must increase by $133^3 \approx 2.35 \text million$ times. $200g \times 2,350,000 = 470,000 kg$.
Conclusion: You cannot accurately sample coarse material with small masses. This explains why "scoop sampling" of conveyors is fundamentally flawed without proper mass reduction protocols (riffle splitters, rotary dividers).
PLS is ideal when you have many collinear predictors (e.g., XRF elemental intensities) and want to predict an assayed grade. PLS finds latent variables that maximize covariance between predictors and responses.
Case: Online XRF analyzers produce raw counts for 15 elements. A PLS model predicts Cu, Zn, and Pb grades with an R² > 0.9 using only spectral data, without needing extensive matrix corrections.
Once significant factors are identified, RSM (e.g., Central Composite Design, Box-Behnken) models curvature. This is essential for finding true maxima (recovery) or minima (cost, reagent consumption).
Output: A contour plot showing predicted recovery vs. two continuous variables, with a clear stationary point.