Snee & Marquardt (1976): Screening Designs for Mixture Experiments

Introduction

In their landmark 1976 Technometrics paper, Ronald D. Snee and Donald W. Marquardt introduced a structured screening design for mixture experiments. Mixture experiments differ from standard factorial or response surface designs because the independent variables are proportions of components that must always sum to one.

Illustration of a three-component mixture design where the proportions of A, B, and C add up to one, a key principle in mixture Design of Experiments.

This inherent constraint requires specialist designs, such as simplex-centroid or simplex-lattice schemes, adapted to deal with the geometry of the feasible region.

Snee and Marquardt’s contribution was twofold:

A systematic way to screen a large number of mixture components with a relatively small set of runs.
A demonstration of how this approach could reduce experimental complexity, identify redundant factors, and accelerate product development programmes.

Mixture Experiment Basics

For a mixture with $q$ components, let the component proportions be:

$x_1, x_2, \dots, x_q$

subject to the conditions:

$x_i \geq 0 \quad \text{for all } i$

$\sum_{i=1}^{q} x_i = 1$

The simplest model for a mixture experiment is the linear mixture model, written without an intercept term as:

$y = \beta_1 x_1 + \beta_2 x_2 + \cdots + \beta_q x_q$

where $y$ is the response and $\beta_i$ are the coefficients representing the influence of each component.

The Snee & Marquardt Screening Design

For effective screening, they recommended a $(3q+1)$ point design. The structure is as follows:

Pure components (each vertex of the simplex):
$(1,0,0,\dots), (0,1,0,\dots), \dots, (0,0,\dots,1)$
Interior mid-points (halfway between a pure component and the centroid).
Overall centroid (all components equal):
$\left(\tfrac{1}{q}, \tfrac{1}{q}, \dots, \tfrac{1}{q}\right)$
Constraint-plane endpoints (points like $(0, \tfrac{1}{q-1}, \dots, \tfrac{1}{q-1})$ and permutations thereof).

The total number of required runs is:

$N = 3q + 1$

This gives adequate information to fit a linear mixture model, while also offering diagnostic capability via response trace plots — graphs that track predicted response as one component is varied from its pure point to the centroid.

The Eight-Component Case Study

In their published example, Snee and Marquardt applied this design to an eight-component mixture. With $q = 8$ , the screening design required:

$N = 3(8) + 1 = 25$

runs in its full form, though they also illustrated a reduced 16-run version for efficiency.

Key Results

Negligible component:
Component $x_6$ produced a flat response trace, demonstrating almost no effect on the outcome. It could therefore be dropped from further study.
Interchangeable components:
The response traces for several pairs or groups of components were nearly identical:
- $x_2 \approx x_3$
- $x_1 \approx x_4$
- $x_5 \approx x_7 \approx x_8$
Model simplification:
This grouping enabled a dimensionality reduction. By excluding $x_6$ and collapsing the similar components, the eight-component system could be re-expressed as an effective three-component mixture.
$x_1^{\prime}=\frac{x_2+x_3}{1-x_6}$
$x_2^{\prime}=\frac{x_1+x_4}{1-x_6}$
$x_3^{\prime}=\frac{x_5+x_7+x_8}{1-x_6}$

This transformation preserved the total sum-to-one condition, whilst eliminating redundancy.

Programme Benefits

The practical advantages of this screening experiment were significant:

Elimination of an inactive component
Component $x_6$ was shown to be unnecessary, saving raw material costs and simplifying formulation decisions.
Grouping of interchangeable components
Components with nearly identical effects could be combined into pooled variables, reducing experimental dimensionality.
More efficient optimisation
The system was reduced from 8 variables to 3 effective mixture variables, making subsequent optimisation designs (e.g. simplex-centroid or response surface refinements) more manageable and less resource-intensive.
Model parsimony
A linear mixture model remained adequate, avoiding the complexity of higher-order polynomials while still providing clear predictive capability.
Strategic focus
Development resources could be concentrated on the components and combinations that truly influenced performance, streamlining the entire programme.

Summary Table of Insights

Finding	Interpretation
$x_6$ had negligible effect	Eliminated from further consideration
$x_2 \approx x_3$	Treated as a single group
$x_1 \approx x_4$	Treated as a single group
$x_5 \approx x_7 \approx x_8$	Treated as a single group
8 components collapsed into 3	Dimensionality reduction
16–25 runs sufficient	Efficient for initial screening

Conclusion

Snee and Marquardt’s 1976 experiment remains a classic case study in the field of mixture Design of Experiments (DoE). By intelligently structuring their screening design, they demonstrated how a high-dimensional mixture could be reduced to its essential factors with only a modest number of runs. The approach directly benefited the programme they were working on by cutting unnecessary complexity, clarifying which components mattered, and providing a clear path forward for optimisation.

This work laid the foundation for modern mixture design strategies, and it continues to influence chemical, food, and materials engineers who must balance efficiency with insight in experimental planning.

Snee & Marquardt (1976): Screening Designs for Mixture Experiments

Introduction

Mixture Experiment Basics

The Snee & Marquardt Screening Design

The Eight-Component Case Study

Key Results

Programme Benefits

Summary Table of Insights

Conclusion

Share this:

Like this:

Leave a ReplyCancel reply

Discover more from Product Development Engineers Ltd