Chapter 28 · Calibration and Flowsheet Simulation

Predictive uncertainty, verification and scale-up

A numerical result becomes an engineering prediction when the calculation, material definition and intended decision are linked to independent evidence. Uncertainty measures how the output can change as evidence changes. Verification checks the implementation of the equations and numerical procedures. Validation asks whether those equations reproduce behavior for the stated use and conditions. These tasks start before calibration, because an untraceable feed assay or an undefined product basis cannot be made specific by a better optimizer. Sargent, 2013

Reasons a prediction may move

Repeat sampling provides information on random measurement uncertainty. Calibration coefficients carry estimation uncertainty. Operating variability describes actual differences among ore blends, feed rates or availability histories. Model-form discrepancy arises when similar operating states give differences the governing equations do not represent. Numerical error includes finite class meshes, solver tolerances and time steps. Keep these terms distinguishable. An expanded interval for pump discharge caused by changing ore is useful for operating planning. A narrow uncertainty band from many replicate pressure readings may still leave a material-flow model weakly validated.

Use plausible dependence in the uncertainty model. Grinding rates at several sizes estimated from one mill survey may share a covariance. Solids and water flows inferred from one slurry reading share inputs. Assign uncertainty to independent measurements or preserve the joint covariance of derived quantities. For uncertain phase fractions, use physically constrained samples in a simplex or fit a covariance on independent coordinates and work within the feasible region. Independently perturbing each normalized fraction and clipping negative values biases the result.

Propagation through an engineering response

Let y = g(x,θ), where x is the operating state and θ is the fitted model parameter vector. For small uncertainties and a differentiable response, define z = (x,θ). If Jg is the Jacobian with respect to z and Σz is their joint covariance, then

Cov(y) ≈ Jg Σz Jgᵀ

When measurement and fitted parameters are correlated, the off-diagonal blocks contribute to this expression. The linearization can be checked by perturbing inputs over their actual uncertainty range. A derivative estimated with a step below the numerical solver tolerance is unreliable. Choose a perturbation that is small physically and large enough to distinguish numerical variation. When bounds, branch changes or long recycles dominate, draw from the joint input distribution and apply the original solver. Keep an explicit count of infeasible or nonconverged cases. Dropping them conditional on a satisfactory answer changes the population being characterized. JCGM 100:2008 JCGM 101:2008

Worked example: uncertainty in metallurgical recovery

For dry feed mass flow F and concentrate mass flow C, let f and c be their valuable-element assays on the same dry basis. R = Cc/(Ff). The feed is 100 ± 3 t/h at 2.00 ± 0.10 g/t. The concentrate is 5.00 ± 0.25 t/h at 32.0 ± 2.0 g/t. Each ± value denotes one standard uncertainty. Assume independent measurements for this illustration. The central feed metal flow is 200 g/h, the concentrate metal flow is 160 g/h and R = 0.800. Differentiation gives

Var(R)/R² ≈ (σC/C)² + (σc/c)² + (σF/F)² + (σf/f)²

The relative standard uncertainty is (0.05² + 0.0625² + 0.03² + 0.05²)¹ᐟ² = 0.0990. The standard uncertainty in R is 0.0792, or 7.92 percentage points. An approximate normal interval at 95% is 0.800 ± 1.96 × 0.0792, or approximately 0.645 to 0.955. This conditional interval addresses the four measurements alone. A common positive relative calibration error in both assays would cancel in their ratio. Its cancellation is represented by covariance across assays. If uncertainty puts appreciable probability near 100% recovery, enforce the physical component balance and analyze the constrained distribution rather than truncating a normal interval as the final estimate.

Verification before comparison with plant data

Build several tests at increasing levels of coupling. A one-class product split with fixed partition has a direct target for each phase. Apply a class-remapping test to conserve the sum of retained masses and check overlap at a mesh boundary. Run a splitter with zero bypass and equal recoveries to exercise obvious limits. For a well-mixed tank with constant flow, compare to the analytic exponential. For an event-driven transfer, force a simultaneous arrival and withdrawal and test the chosen order. Cross-check each quantity in independent units when code has a conversion boundary, especially kPa versus Pa, µm versus mm and t/h versus kg/s.

Numerical verification must include a convergence study where an equation has been discretized. Narrow the class mesh where a partition turns sharply. Compare product mass and representative P80 over at least two suitable meshes. Shorten the integration interval at fixed physical parameters and compare total accumulation, peak levels and event timing. Choose refinement based on process scales and transport delay. A solid or water loss revealed at a mesh change is a problem to solve even if an empirical refit would restore P80.

Validation across scale and operating domain

Model validation begins with an explicit acceptance statement. The following matrix illustrates a plant grinding study rather than universal tolerances.

OutputSuggested evidence for the studyFailure decision
Total dry grinding throughputIndependent belt measurement over a stable survey; accountable external boundarySuspend comparison if the rate does not close within the survey uncertainty
Product PSDSize samples under several feed and water conditions; method for censoring the finest intervalTest whether error concentrates near classifier cut or in a mill appearance region
Circulating loadMeasured recycle or several redundant mass/sizing balancesFlag errors in the recycle branch separately from fresh product tonnage
Net powerMetered active power and measured idle basis where relevantDistinguish power entering the process from gross draw
Buffer peak levelTransient survey or conservation test around a known disturbanceInspect initial fill, delay, overflow logic and measurement cadence

Published HPGR work illustrates the significance of equipment and scale. A simple bulk strip expression for an illustrative capacity is Q = 3600 ρb s W u C, where dry mass density ρb is in t/m³, working opening s and roll width W are in m, representative bed velocity u is in m/s, and C is a dimensionless effective fill factor. This illustrates geometry-based throughput scaling and omits material breakage and edge information. A pilot with ρb = 2.1 t/m³, s = 0.012 m, W = 0.25 m, u = 0.70 m/s and C = 0.8 yields 12.701 t/h. Using 0.035 m, 1.4 m and 1.7 m/s in the corresponding plant expression gives 503.798 t/h. An assumed effective fill change from 0.8 to 0.72 lowers the estimate by 10%. That shift should be explained by evidence such as geometry, feed or an edge effect, with checks of pressure, actual gap and product PSD. In HPGR model validation, Daniel and Morrell reported sensitivity of throughput and estimated power to inconsistent gap-to-diameter ratios across scales. Daniel and Morrell, HPGR model verification and scale-up

A useful sign-off records the calibrated domain, the independently tested domain, the inputs that move the decision and the outputs still conditional on extrapolation. Repeat or redesign the experiment when an influential assumption remains weakly supported.

Exercises and answers

  1. If the same multiplicative calibration factor affects both feed and concentrate assays, what does it do to R = Cc/(Ff)? It cancels algebraically where the factor is genuinely shared. Independent sample preparation errors remain.
  2. A second HPGR design retains the same bulk density, fill factor, gap and speed but doubles roll width. What strip-model change follows? Estimated throughput doubles. Real throughput needs evidence about effective fill and edge behavior at the new width.
  3. Two grind simulations have closed mass balances and the same P80 but coarse tails differ by 6 percentage points. What output should be validated if the decision is trash-screen capacity? The relevant coarse product size classes and their mass flows. P80 is one percentile and cannot establish those tail flows.