Chapter 27 · Calibration and Flowsheet Simulation

Calibration, identifiability and model selection

Calibration estimates model parameters from observations under documented operating conditions. A useful estimate must retain a physical interpretation, reproduce more than one observable when the model claims to do so, and support the prediction needed by the study. A mathematically successful fit can leave individual parameters uncertain or depend strongly on one operating campaign.

Choose measurements that distinguish mechanisms

Separate direct inputs from fit parameters. Measure geometry, opening dimensions, slurry properties or rotor speed when a reliable physical measurement exists. Use the measured geometry in the correlation and assign its uncertainty in later propagation. Reserve fitted multipliers for variation the correlations leave unresolved. A separate empty-equipment power test may identify no-load power more clearly than a product size distribution can. A water split in a sampled classifier may help separate a fines bypass from genuine near-cut selectivity.

A good calibration set contains variation in feed, duty or operating controls that breaks correlations between parameters. Samples taken at six almost identical cyclone pressures improve estimates of repeatability but provide little leverage for the pressure exponent. An operating range with measured changes in cut and recovery can do more. Include extreme fractions only when the instrumentation resolves them. An apparent plateau in the finest class can represent a real bypass, sample contamination or the resolution limit of an assay.

State the fitted target. A fit to PSD normalized to 100% says little about product tonnage without a mass-flow observation. Two curves with opposite mass splits can have the same normalized product PSD. When fitting both streams, use either absolute class flow rates with their covariance or mass flows and independent PSD coordinates. A check of both wet mass and calculated dry and water mass from that same flow reading is useful for accounting but does not supply independent measurements.

Objective and sensitivities

Write the prediction vector f(θ) for parameter vector θ, and let the observation vector be y with covariance Σ. Consider

Φ(θ) = (f(θ) − y)ᵀΣ⁻¹(f(θ) − y)

If a Cholesky factor C satisfies Σ = CCᵀ, the whitened residual is r = C⁻¹(f − y). Its Jacobian J has entries measuring the changes of whitened predictions with respect to parameters. For a differentiable unconstrained iterate, linearize r(θ + Δ) ≈ r + JΔ and solve the least-squares problem for Δ. Use a QR or singular-value factorization when sensitivity columns are nearly dependent. Trust-region and damping methods restrict steps when the linearization is poor. Bounds may encode physically sensible parameter signs, but a coefficient pinned at a bound deserves inspection, particularly if predictions at unsampled duties depend on it.

For constrained optimization and many empirical fits, work in coordinates suited to the physics. A logarithmic parameter maps positive values onto the real line, while a logistic transformation keeps a recovery fraction between zero and one. The associated Jacobian transforms the uncertainty back to physical coordinates. Always report the actual fit function and the weighting basis. Taking logarithms of measurements changes the residual model and emphasizes relative rather than absolute differences.

Interpret residuals graphically by measurement order, size, operating condition and component. An orderly pattern at fine and coarse sizes can indicate a curve shape the implementation cannot express. One oversized residual during an outage can reflect a different state rather than a curve coefficient. Large residuals in low-flow streams can arise from missing or poorly sampled solids. A fitted multiphase system often benefits from a staged strategy. Fit flow or power first where its equation has a distinct observation. Fit the size or component split against the measured phase flows. Finish with a joint consistency check so that cross-coupling is retained.

Structural and practical identifiability

A useful screen-passage example has probability P(d) = 1 − exp[−nκ(d)], where n is an effective number of trials and κ(d) describes passage per trial under the assumed model. If κ(d) = a g(d) and both n and a are treated as free, every prediction depends only on the product n a. For any c > 0, parameters cn and a/c produce the same curve. This is structural non-identifiability. Fix one group or measure it independently. Merely fitting more points on the same curve cannot remove the exact parameter equivalence.

Practical non-identifiability occurs when the observations weakly resolve parameters that are theoretically separable. A smooth fit with a broad family of possible sharpness coefficients can arise if the test samples cover mostly fine or mostly coarse sizes. In the whitened sensitivity matrix, nearly parallel columns appear as a small singular value. Profile a parameter by fixing it at a series of values and reoptimizing the rest. A relatively flat profile warns that a confidence bound from local quadratic curvature could be misleading. For regular interior parameters with a well-specified likelihood, the profile-likelihood method can expose asymmetric intervals or open bounds. Raue et al., 2009

Curvature-based standard errors are conditional. With known variance, an interior full-rank approximation is Cov(θ̂) ≈ (JᵀJ)⁻¹. If the variance scale was estimated from the same fit using independent errors, include the corresponding residual scale. Never interpret this inverse as evidence of model-form correctness. Predictions can be better identified than the individual parameters. Report uncertainty in the outputs needed for the decision.

Worked example: separating bypass and sharpness

Consider this illustrative actual coarse partition, with independent water recovery giving B = 0.20 and a corrected logistic curve C(d).

C(d) = 1 / [1 + (d₅₀c/d)ᵃ]; R(d) = B + (1 − B) C(d)
Particle size d (µm)Actual Coarse partition RCorrected C = (R − 0.20)/0.80
250.247060.05882
500.360000.20000
1000.600000.50000
2000.840000.80000

At d = 100 µm the corrected probability is one-half, so d₅₀c = 100 µm. Writing ln[C/(1 − C)] = a ln(d/d₅₀c) gives −ln 4 = a ln 0.5 at 50 µm and ln 4 = a ln 2 at 200 µm. Both give a = 2. In this constructed case the observations agree exactly. For real data, fit R on its original measurement basis with appropriate weighting. Points near physical bounds can give a poor residual model after a logit transformation.

At d₅₀c the actual coarse partition is 0.60, since the bypass adds 0.20 to 0.80 × 0.50. If the 0.50 point on the actual curve were used to define the model cut, the value would be lower. The actual 0.50 point occurs where C = 0.375, or d = 100 √(0.375/0.625) ≈ 77.46 µm. A partition calculation may therefore provide two valid cut conventions with different meanings. The use of water recovery as B here is an explicit model assumption, to be checked against fine solids observations before extending to other feed conditions.

Selecting a model for prediction

Prefer a structure whose assumptions can be tested with available observations and whose prediction survives independent operating conditions. Compare candidate models on the same response vector, data subset and likelihood. AIC = 2k − 2 ln L̂ encourages predictive parsimony within such a comparison, with k defined as the number of independently fitted parameters and L̂ the maximum likelihood. AIC differences are invalid when one candidate fits mass flows and another fits only PSD percentages with a different observation set. Additional physics and measured observables can matter more for extrapolation than a small improvement in fit. Akaike, 1974

For plant surveys with repeat measurements, split validation by day, ore domain, operating campaign or unit. Randomly separating individual size classes from one normalized PSD leaks information from the same sample. A final calibration note should state the parameter set, bounds, data used, residual assumptions, fit quality by observable, validation range and plausible failure modes. A third party should be able to reconstruct the analysis from its raw measurements and study boundary.

Exercises and answers

  1. Can retention data at ten particle sizes identify n and a separately if retention is exp[−n a g(d)]? No. The data identify at most the product n a. Add an independent measurement or fix one factor using a justified physical basis.
  2. If a corrected cut remains fixed and B rises from 0.10 to 0.30, what is the change in actual recovery at the corrected cut? At that size R = 0.5 + 0.5B. The actual probability changes from 0.55 to 0.65.
  3. Two models have the same fitted likelihood on identical observations. Model A uses three independently fitted parameters and Model B five. What is AIC_B − AIC_A? Four. A is favored on this criterion. Consider the predicted outputs separately if the models differ in their validity range.