6SQ ToolsWhat Is a D-Optimal Mixture Design and When Should You Use It?
If you formulate products—paint, concrete, food blends, or chemical mixtures—you already know the challenge: the response depends on the proportions of ingredients, not the total amount. A standard factorial design won't work because the factors are not independent. That is where a D-optimal mixture design comes in. This article explains what it is, how it works, and how to run one using a free tool.
What It Is
A mixture design is a class of experimental designs used when the factors are components of a mixture, and their sum is fixed (e.g., 100%). The response is a function of the proportions of each component, not the absolute quantities.
A D-optimal design is a computer-generated design that selects a subset of candidate points (from a simplex region) to maximize the determinant of the information matrix (X'X). In plain terms, it chooses the set of runs that gives the most precise estimates of the model coefficients, given a specified model (linear, quadratic, or special cubic) and a fixed number of runs.
The theory is well established in the classic reference Experiments with Mixtures by John Cornell. D-optimality is a standard optimality criterion used when the classical simplex-lattice or simplex-centroid designs are impractical—for example, when you have constraints on component ranges or when you need fewer runs than a full design.
How It Works / Formula or Steps
Each component \(x_i\) must satisfy \(0 \le x_i \le 1\) and \(\sum x_i = 1\). You may also have upper/lower bounds (e.g., \(0.2 \le x_1 \le 0.6\)).
Common models are:
- Linear: \(y = \sum \beta_i x_i\)
- Quadratic: \(y = \sum \beta_i x_i + \sum_{i<j} \beta_{ij} x_i x_j\)
- Special cubic: adds terms like \(\beta_{ijk} x_i x_j x_k\)
These are points on the simplex vertices, edges, and interior (from simplex-lattice or simplex-centroid designs) that satisfy your constraints.
From the candidate set, the algorithm selects \(n\) runs to maximize \(\det(X'X)\), where \(X\) is the model matrix. A larger determinant means smaller confidence ellipsoids for the coefficients—i.e., more precise estimates.
Analyze the data with the appropriate mixture model, check for lack of fit, and optimize the formulation.
A Worked Illustrative Example
Example data (illustrative only):
Suppose you want to optimize a three-component blend (A, B, C) with a quadratic model. You have constraints: A between 0.2 and 0.6, B between 0.1 and 0.5, C between 0.1 and 0.4, and A + B + C = 1.
A full simplex-lattice design might require many runs. Instead, you use a D-optimal design with 10 runs. The algorithm generates candidate points (vertices, edge midpoints, and centroid) and selects 10 that maximize \(\det(X'X)\).
The selected design might look like this:
Run | A | B | C
After running the experiments and measuring the response (e.g., viscosity), you fit the quadratic mixture model and identify the optimal blend. The D-optimal design ensures that your coefficient estimates are as precise as possible with only 10 runs.
Common Pitfalls
Try It Free
You don’t need expensive software to generate a D-optimal mixture design. Use the free tool at https://www.6sq.com/tools/mixture_dopt/ to enter your components, constraints, and model, and get a ready-to-run design in seconds.
What It Is
A mixture design is a class of experimental designs used when the factors are components of a mixture, and their sum is fixed (e.g., 100%). The response is a function of the proportions of each component, not the absolute quantities.
A D-optimal design is a computer-generated design that selects a subset of candidate points (from a simplex region) to maximize the determinant of the information matrix (X'X). In plain terms, it chooses the set of runs that gives the most precise estimates of the model coefficients, given a specified model (linear, quadratic, or special cubic) and a fixed number of runs.
The theory is well established in the classic reference Experiments with Mixtures by John Cornell. D-optimality is a standard optimality criterion used when the classical simplex-lattice or simplex-centroid designs are impractical—for example, when you have constraints on component ranges or when you need fewer runs than a full design.
How It Works / Formula or Steps
- Define the components and constraints.
Each component \(x_i\) must satisfy \(0 \le x_i \le 1\) and \(\sum x_i = 1\). You may also have upper/lower bounds (e.g., \(0.2 \le x_1 \le 0.6\)).
- Choose a model.
Common models are:
- Linear: \(y = \sum \beta_i x_i\)
- Quadratic: \(y = \sum \beta_i x_i + \sum_{i<j} \beta_{ij} x_i x_j\)
- Special cubic: adds terms like \(\beta_{ijk} x_i x_j x_k\)
- Generate candidate points.
These are points on the simplex vertices, edges, and interior (from simplex-lattice or simplex-centroid designs) that satisfy your constraints.
- Apply the D-optimality criterion.
From the candidate set, the algorithm selects \(n\) runs to maximize \(\det(X'X)\), where \(X\) is the model matrix. A larger determinant means smaller confidence ellipsoids for the coefficients—i.e., more precise estimates.
- Run the experiment and fit the model.
Analyze the data with the appropriate mixture model, check for lack of fit, and optimize the formulation.
A Worked Illustrative Example
Example data (illustrative only):
Suppose you want to optimize a three-component blend (A, B, C) with a quadratic model. You have constraints: A between 0.2 and 0.6, B between 0.1 and 0.5, C between 0.1 and 0.4, and A + B + C = 1.
A full simplex-lattice design might require many runs. Instead, you use a D-optimal design with 10 runs. The algorithm generates candidate points (vertices, edge midpoints, and centroid) and selects 10 that maximize \(\det(X'X)\).
The selected design might look like this:
Run | A | B | C
- 1 | 0.60 | 0.30 | 0.10
- 2 | 0.20 | 0.50 | 0.30
- 3 | 0.50 | 0.10 | 0.40
- ... | ... | ... | ...
- 10 | 0.35 | 0.35 | 0.30
After running the experiments and measuring the response (e.g., viscosity), you fit the quadratic mixture model and identify the optimal blend. The D-optimal design ensures that your coefficient estimates are as precise as possible with only 10 runs.
Common Pitfalls
- Using D-optimal when a standard design works. If you have no constraints and a simple model, a simplex-lattice or simplex-centroid design is simpler and easier to interpret.
- Choosing too few runs. D-optimal designs can be very efficient, but if you select fewer runs than the number of model terms, you cannot estimate all coefficients.
- Ignoring the mixture constraint. Always verify that your component proportions sum to 1 in every run.
- Overfitting with a high-order model. A special cubic model needs many runs; use it only when theory supports it.
Try It Free
You don’t need expensive software to generate a D-optimal mixture design. Use the free tool at https://www.6sq.com/tools/mixture_dopt/ to enter your components, constraints, and model, and get a ready-to-run design in seconds.
No related results found
Invited:
0 replies