Examples / 03

Mixture with five components and constraints: an emulsion paint

In a formulation no component can be changed without another giving way: the shares add up to 100%. Add bounds and ratio requirements, and what is left of the mixture space is an irregular body for which there is no standard design. This example shows the way from five components and four kinds of constraints to a recipe that meets three responses at once.

Design
I-optimal mixture design
Runs
16 + 5 + 4 = 25
Components
5, with bounds, a sum and a ratio constraint
Responses
Viscosity, hiding power, wet scrub loss

The measurements are simulated. They come from three models we set ourselves, plus random scatter — which makes it possible to check at the end whether the analysis finds what is really in there. Every table and chart was computed and drawn by DoEStat.

Step 1 / The question

Five components that add up to 100%

The paint consists of binder, pigment, filler, water and an additive package. For each component there is a lower and an upper bound — below the one the paint does not work, above the other it becomes too expensive or can no longer be processed:

ComponentLower bound [%]Upper bound [%]Actually in the design [%]
Binder305030.0 … 50.0
Pigment152515.0 … 25.0
Filler5255.0 … 25.0
Water103010.0 … 30.0
Additive262.0 … 6.0

In addition there are two conditions that link several components:

  • Pigment + filler ≤ 40% — the dispersion no longer binds more solids reliably.
  • Binder : pigment ≥ 1.6 — every part of pigment needs at least 1.6 parts of binder.

Three quantities are measured: the viscosity in Krebs units (target 100 KU, 90 to 110 allowed), the hiding power as contrast ratio (to be maximised, at least 96%, ideally 98%) and the wet scrub loss in micrometres (to be minimised, at most 20 µm, ideally 8 µm). The wet scrub loss counts somewhat less in the trade-off than the other two.

Step 2 / The region

What the constraints leave of the mixture space

Five shares adding up to one span a four-dimensional body. Without bounds it would be a regular simplex, and the classical simplex designs would fit. The lower and upper bounds cut a body with 24 vertices out of it; the sum and the ratio constraint cut once more and leave 28 vertices. DoEStat computes this body from the constraints as entered — the ratio constraint is written as what it is, a straight line: binder − 1.6 · pigment ≥ 0.

The vertices, the midpoints of edges and faces and the centroid give 124 candidate points. Running all of them would be a waste: a quadratic mixture model for five components has 15 coefficients.

Step 3 / Design

16 points for the model, 9 for the check

An optimal design selects from the candidates the points with which the desired model can be estimated best. The I-criterion is chosen: it minimises the average prediction variance over the whole feasible region — the right thing when a recipe is to be predicted at the end. 16 points carry the model; added to these are five points that DoEStat places in the gaps to expose lack of fit, and four replicates for pure error.

RunPointBinder [%]Pigment [%]Filler [%]Water [%]Additive [%]Viscosity [KU]Hiding power [%]Wet scrub loss [µm]
1edge midpoint48.020.020.010.02.0110.797.1210.1
2lack-of-fit34.421.512.130.02.032.296.3217.7
3vertex40.025.015.014.06.0148.197.9614.6
4edge midpoint50.020.05.023.02.046.695.235.4
5lack-of-fit30.015.025.024.06.093.094.5027.8
6replicate41.119.714.021.14.083.895.9713.4
7replicate44.120.55.024.46.075.895.3210.2
8edge midpoint39.015.014.030.02.030.893.6616.5
9lack-of-fit42.617.419.515.65.0114.495.7213.3
10centroid41.119.714.021.14.083.095.8615.2
11lack-of-fit35.021.918.119.06.0122.897.4821.4
12edge midpoint37.523.55.030.04.040.995.7917.9
13edge midpoint40.025.010.023.02.056.296.5014.4
14replicate50.025.08.013.04.0114.597.315.7
15face centroid50.025.08.013.04.0113.897.286.9
16vertex30.015.019.030.06.058.494.0622.5
17face centroid44.120.55.024.46.075.195.5510.6
18lack-of-fit44.015.025.010.06.0151.595.7716.7
19centroid41.119.714.021.14.082.095.8812.1
20face centroid48.015.05.028.04.040.692.885.8
21vertex30.018.721.328.02.047.895.9725.1
22face centroid38.015.025.018.04.098.994.9519.4
23vertex50.015.019.010.06.0151.195.388.1
24replicate41.119.714.021.14.082.896.4212.5
25edge midpoint50.015.014.019.02.058.493.606.7

Among the 16 model points are four vertices, six edge midpoints, four face centroids and, twice, the centroid: an optimal design may repeat a point when that improves the prediction. With the four replicates the design thus has five degrees of freedom for pure error. The last column of the bounds table above shows that the design moves every component across its whole allowed range — as far as the two linking constraints permit.

Step 4 / Models

Scheffé models: linear, and with blending terms where needed

Mixture models have no constant — it is contained in the linear terms, because the shares add up to one. The linear terms describe how each component acts on its own; the blending terms (products of two components) describe where two components together do more or less than the sum.

ResponseModel orderSequential pLack of fit pAdj. R²Pred. R²
ViscosityLinear< 0.0001< 0.00010.98350.9755
ViscosityQuadratic0.00520.00100.99510.8978
Hiding powerLinear< 0.00010.09610.92100.8943
Hiding powerQuadratic0.00850.59390.97400.9268
Wet scrub lossLinear< 0.00010.31410.95010.9362
Wet scrub lossQuadratic0.38350.30140.95490.6636

The table is a warning. For all three responses the full quadratic model has the higher adjusted R² — but the higher predicted R² only for the hiding power; for the viscosity it drops to 0.90, for the wet scrub loss as low as 0.66. Ten blending terms on 25 runs are too many: the model follows the noise. The answer lies in between. DoEStat starts from the quadratic model, works backwards and keeps only the blending terms that are significant; the linear terms always stay.

ResponseBlending terms in the modelR²Adj. R²Pred. R²Residual sLack of fit p
Viscosity [KU]30.99750.99640.99262.200.0035
Hiding power [%]40.97850.96770.94710.2290.4597
Wet scrub loss [µm]10.96850.96020.94441.230.4443

One to four blending terms per response are enough. The predicted R² is now above 0.94 for all three. For the hiding power and the wet scrub loss the lack of fit is not significant; for the viscosity the test fires (p = 0.0035). It compares what the model leaves unexplained with the scatter between identical batches — and here that scatter, at 0.65 KU from five degrees of freedom, is strikingly small next to the model’s residual scatter of 2.2 KU. Whether the model is missing something or the replicates happened to lie close together, the test cannot tell; the confirmation and the cross-check at the end show which it is.

Term (pseudo-components)Viscosity [KU]Hiding power [%]Wet scrub loss [µm]
Binder87.497.536.1
Pigment9.1965.519.8
Filler24.493.76.53
Water2621017.04
Additive−1571045.26
Binder × Pigment·28.8·
Pigment × Filler10533.6−42.2
Pigment × Water−55.725.2·
Water × Additive−439−17.2·

The coefficients are in pseudo-components: every component is rescaled so that 0 is its lower bound. A linear coefficient is then the prediction for the recipe in which all other components sit at their lower bounds and this one gets the whole remainder. Because every component here also has an upper bound, these recipes lie outside the region studied: the linear coefficients are extrapolations and cannot be read one by one — in real shares it would even be the prediction for “100% additive”. The predictions themselves are the same in both notations.

Analyses of variance of the three models

In the analysis of variance of a mixture the linear terms appear as one line: asking whether a single linear coefficient is zero would be asking whether the pure component has a response of zero. The sensible question is whether the response depends on the recipe at all.

SourceSum of squaresdfMean squareFp
Model3254274649961.67< 0.0001
Linear mixture32176480441663.97< 0.0001
Pigment × Filler52.37152.3710.830.0043
Pigment × Water23.96123.964.960.0398
Water × Additive287.91287.959.56< 0.0001
Residual82.18174.834
Lack of fit80.05126.67115.660.0035
Pure error2.13050.4260
Total3262524
SourceSum of squaresdfMean squareFp
Model38.0184.75190.86< 0.0001
Linear mixture36.2949.073173.50< 0.0001
Binder × Pigment0.932010.932017.820.0006
Pigment × Filler1.48111.48128.32< 0.0001
Pigment × Water1.19411.19422.840.0002
Water × Additive0.348210.34826.660.0201
Residual0.8367160.05229
Lack of fit0.6027110.054791.170.4597
Pure error0.234050.04680
Total38.8524
SourceSum of squaresdfMean squareFp
Model882.05176.4116.76< 0.0001
Linear mixture872.94218.2144.43< 0.0001
Pigment × Filler9.17719.1776.070.0234
Residual28.71191.511
Lack of fit22.21141.5861.220.4443
Pure error6.50051.300
Total910.724
Residuals of the viscosity against predicted values: the points scatter around the zero line without a pattern.
Residuals against predictionViscosity: even scatter from 28 to 154 KU.
Normal quantile plot of the standardised residuals of the viscosity: the points follow the straight line.
Normality of the residualsThe points follow the line.

Step 5 / Component effects

Which component moves which response

In a mixture there is no “change this component and hold everything else”. What one looks at instead: start at a reference recipe, move one component through its allowed range and let the others give way in the same proportion to one another. The table shows the change of each response along that path, starting from the centroid of the region.

ComponentViscosity [KU]Hiding power [%]Wet scrub loss [µm]
Binder3.68−1.37−17.4
Pigment20.33.322.23
Filler36.80.80810.0
Water−110−3.005.42
Additive41.80.01381.00

The roles are clearly divided. The viscosity depends on the water (−110 KU across its range) and on the additive, which moves 42 KU despite its small amount. The hiding power depends on the pigment and, in the opposite direction, on the water; the wet scrub loss on the binder. The trace plots show the same as curves, here starting from the optimal recipe:

Trace plot of the viscosity: five lines, one per component, through a common point at 100 KU. The line for water falls steeply, that for additive rises steeply, filler and pigment rise gently, binder is almost level.
Trace: viscosityWater thins, additive thickens. At the optimum the additive is at its lower bound — its line can only go to the right.
Trace plot of the hiding power: the line for pigment rises and flattens, that for water falls clearly, filler and additive rise slightly, binder falls slightly.
Trace: hiding powerMore pigment hides better, but with diminishing returns — the curve flattens. At the optimum the pigment sits at its upper bound; its line can only go left.
Trace plot of the wet scrub loss: the line for binder falls steeply, those for filler and water rise, pigment rises more gently, additive falls slightly.
Trace: wet scrub lossBinder holds the film together; filler and water in its place weaken it.

For surface pictures a mixture needs the triangle. Five components do not fit into it, so two are held — here at the values of the optimal recipe — and three share the rest among themselves. Grey is what the bounds and the two linking constraints exclude.

Ternary contour plot of the viscosity over binder, filler and water: a coloured quadrilateral in the otherwise grey triangle, the contour lines run almost parallel, the values fall towards the water corner from about 110 to 35 KU.
Viscosity in the trianglePigment and additive held. The viscosity falls evenly towards the water corner.
Ternary contour plot of the wet scrub loss over binder, filler and water: the smallest values lie on the binder side of the feasible area.
Wet scrub loss in the triangleThe same cut: the loss is smallest where there is much binder and little filler.
Ternary contour plot of the hiding power over binder, pigment and filler: the feasible area is a narrow coloured field, the values rise towards the pigment corner.
Hiding power in the triangleWater and additive held, pigment on an axis. The sum constraint bounds the coloured field at the bottom, the upper bound of the pigment on the left.

Step 6 / Optimisation

The best feasible recipe

As in the other examples, desirability combines the three goals into one number. The difference: the search is not in a cube but in the body with 28 vertices — every recipe examined must keep all bounds and both linking constraints and add up to 100%.

ComponentShare [%]Allowed [%]
Binder48.330 … 50
Pigment25.015 … 25
Filler12.35 … 25
Water12.310 … 30
Additive2.02 … 6
Pigment + filler37.3≤ 40
Binder : pigment1.93≥ 1.6
ResponseGoalPrediction95% prediction intervalDesirability dTrue value
Viscosity [KU]target 100 (90.0 … 110)100.094.3 … 105.71.00097.5
Hiding power [%]maximise, at least 96.097.6497.05 … 98.240.82297.44
Wet scrub loss [µm]minimise, at most 20.08.05.1 … 11.00.9987.4
Overall desirability D0.928

The overall desirability is 0.93. The viscosity hits its target exactly. The recipe uses the region to the full: additive at its lower bound, pigment at its upper bound, the solids sum at 37.3% just below the 40% allowed. The ratio of binder to pigment, at 1.93, is well above the minimum — the large amount of binder lowers the wet scrub loss.

All three responses over the triangle

What does the window around this recipe look like? Binder and additive stay at their optimal 48.3% and 2.0%; pigment, filler and water share the rest and span the triangle. Of the ten possible triangles this one leaves the most area — with the binder on an axis, its minimum ratio to pigment leaves at most a scant third. Above the triangle the three responses stand as surfaces, each on its own range from 0 to 1. The grey floor is what can be mixed; yellow on it is what keeps every limit.

Overlaid 3-D surfaces over the mixture triangle of pigment, filler and water with binder and additive fixed: viscosity in blue and hiding power in red fall towards the water corner, wet scrub loss in green rises towards the filler. On the grey floor of the allowed blends a yellow region keeps every limit, along the low-water edge towards the filler corner.
Viscosity, hiding power and wet scrub loss over the triangleBinder 48.3% and additive 2.0% fixed, the triangle in pseudo-components of the remaining 49.7%. “Rotate in 3D” makes the scene turnable; hovering shows recipe and value.

Viscosity and hiding power both fall towards the water corner — water thins the paint, and it displaces pigment. Because both have a lower limit (90 KU, 96%), the window lies along the low-water edge. Wet scrub loss rises towards the filler but stays below its 20 µm everywhere in this cut; it draws no limit here. In DoEStat, the 3D surface page with the Overlaid switch shows this picture for mixtures.

Step 7 / Confirmation

Three batches of the recommended recipe

Response95% prediction intervalConfirmation 1Confirmation 2Confirmation 3Joint test p
Viscosity [KU]94.3 … 105.797.592.8101.60.0467
Hiding power [%]97.05 … 98.2497.3397.6097.160.3330
Wet scrub loss [µm]5.1 … 11.08.79.47.80.7458

Result: hiding power and wet scrub loss confirm the prediction: 97.4% and 8.6 µm on average, all six values within the prediction interval. For the viscosity the second batch, at 92.8 KU, falls below it, and the joint test of the three batches narrowly fires (p = 0.047): the recipe delivers 97.3 KU on average instead of the predicted 100 — still well inside the allowed band of 90 to 110 KU, but the model overestimates the viscosity at this point. All this with 25 trial batches instead of a grid over five components.

Cross-check

What is really in the data

The measurements come from these three models. The letters stand for the deviation of the share from a reference recipe (binder 40%, pigment 20%, filler 15%, water 20%, additive 5%), written as a fraction: b for binder, p for pigment, f for filler, w for water, a for additive.

Viscosity = 100 + 40·b + 220·p + 180·f − 380·w + 1100·a − 3000·w·a + 1200·p·f  (σ = 2.0)
Hiding power = 96.5 − 4·b + 30·p + 6·f − 8·w − 250·p² + 40·p·f  (σ = 0.25)
Wet scrub loss = 14 − 60·b + 25·p + 45·f + 20·w + 300·f² + 150·b·f  (σ = 1.0)

For the viscosity the analysis found the two blending terms that exist — pigment × filler and water × additive — and a third, pigment × water (p = 0.040), that does not exist in truth. For the hiding power the truth lies in a square of the pigment — the saturation the trace plot shows. A Scheffé model knows no squares; because the shares add up to one, p² appears there as the blending terms of the pigment with all the other components. The analysis kept three of those four, pigment × additive is missing; instead, water × additive is in the model (p = 0.020), which does not exist. For the wet scrub loss the decomposition is even less clear-cut: a single blending term was found, pigment × filler, which approximates the true filler square and its interaction with the binder within the region studied but is not the same term.

What counts is the prediction. At the optimal recipe the true values are 97.5 KU, 97.44% and 7.4 µm — predicted were 100.0, 97.64 and 8.0. All three lie within the prediction interval. The model overestimates the viscosity by 2.5 KU; that is exactly what the confirmation noticed. The lack-of-fit test, on the other hand, was a false alarm: the model contains both true blending terms, and the pure error of 0.65 KU lies far below the true σ = 2.0 — the replicates happened to lie close together. Five degrees of freedom are not much for this estimate.

Project files

Run it yourself

The example ships with DoEStat: Help ▸ Open sample project ▸ Mixture ▸ “Five-component mixture with constraints: emulsion paint”, once with the data only and once with the finished analysis. The same files can be downloaded here.

Back: response surface with three responses Next: from screening to a response surface

Try DoEStat for free

30 days, the full feature set, no payment details. We send the download link by e-mail, usually on the next business day.

Request the trial