Mixture Design is a specialized experimental design methodology for optimizing formulations where component proportions must sum to a fixed constant — typically 100% — where the constraint that x₁ + x₂ + ... + xₖ = 1 invalidates standard factorial designs (since components cannot be varied independently), requiring the simplex-based designs and Scheffé polynomial models specifically developed for constrained mixture spaces, with applications spanning CMP slurry formulation, photoresist solvent systems, alloy compositions, and cleaning chemistry optimization.
Why Standard Designs Fail for Mixtures
In a standard two-level factorial design, each factor is varied independently between its low and high values. For a mixture, this is mathematically impossible: increasing component A necessarily decreases at least one other component to maintain the sum = 1 constraint.
Example: Three-component slurry (abrasive particles A, oxidizer B, surfactant C).
- Cannot set A = 0.7, B = 0.7, C = 0.7 (sum = 2.1 ≠ 1)
- Varying A from 0.3 to 0.5 automatically changes B + C by -0.2
The experimental space for a k-component mixture is a (k-1)-dimensional simplex — a triangle for 3 components, tetrahedron for 4, etc.
Standard Mixture Designs
| Design Type | Points Included | Purpose |
|------------|----------------|---------|
| Simplex Lattice {k,m} | All compositions with xᵢ = 0, 1/m, 2/m, ..., 1 | Systematic coverage of simplex |
| Simplex Centroid | Vertices, edge midpoints, face centroids, overall centroid | Balanced exploration, efficient for interactions |
| Extreme Vertices | Vertices of constrained feasible region | When components have min/max bounds |
| D-optimal | Computer-generated, minimizes det(X'X)⁻¹ | Constrained regions, optimal for specific models |
| Augmented Designs | Above + interior points or star points | Better pure error estimation |
Scheffé Polynomial Models
Standard polynomial regression cannot be used for mixtures because of the collinearity induced by the sum constraint. Scheffé (1958) derived reparametrized models:
Linear (first-order): η = Σᵢ βᵢxᵢ (k parameters, no intercept — intercept absorbed into βᵢ)
Quadratic: η = Σᵢ βᵢxᵢ + Σᵢ<ⱼ βᵢⱼxᵢxⱼ (adds pairwise interaction terms)
Special Cubic: Adds βᵢⱼₖxᵢxⱼxₖ terms for three-way interactions
The quadratic model is most commonly used — it captures synergistic and antagonistic blending behavior (βᵢⱼ > 0 indicates synergy: the blend performs better than the linear combination of pure components).
Constrained Mixture Designs
Real formulations impose additional constraints beyond the sum = 1:
- Component lower bounds: xᵢ ≥ Lᵢ (minimum concentration for performance or stability)
- Component upper bounds: xᵢ ≤ Uᵢ (cost, toxicity, or processing constraints)
- Linear inequality constraints: xᵢ + xⱼ ≤ 0.4 (combined concentration limit)
These constraints transform the simplex into an irregular polyhedron. The feasible region's extreme vertices become the natural design points, and D-optimal or I-optimal computer-generated designs are used.
Semiconductor Applications
CMP (Chemical Mechanical Planarization) Slurry Optimization:
Components: Abrasive particles (colloidal silica or ceria), oxidizer (H₂O₂), pH buffer, corrosion inhibitor, surfactant.
Objective: Maximize removal rate for target material while minimizing dishing, erosion, and scratch defects.
Scheffé quadratic model identifies synergistic interactions (e.g., oxidizer + surfactant combination outperforms either alone).
Photoresist Solvent System:
Components: PGMEA (primary solvent), GBL, cyclohexanone.
Objective: Optimize viscosity for spin coating, dissolution contrast, and development rate.
Cleaning Chemistry:
Components: HF, H₂SO₄, H₂O₂, DI water.
Objective: Maximize native oxide removal rate while minimizing silicon loss and metallic contamination.
Analysis and Optimization
After fitting the Scheffé model, optimization uses constrained nonlinear programming to find the component proportions maximizing (or minimizing) the predicted response, subject to the mixture constraints. Desirability functions handle multi-response optimization (simultaneously optimize removal rate AND non-uniformity). The prediction variance across the simplex quantifies confidence in the model predictions for any proposed formulation.