8.7 Response Surface Methodology and Optimization

Response Surface Methodology (RSM) is an advanced collection of mathematical and statistical techniques used in the Improve phase when the objective is to optimize a continuous response $Y$ across continuous control factors $X_i$. While factorial designs identify significant main effects and interactions, RSM models second-order curvature to identify optimal operating points.

---

First-Order vs. Second-Order RSM Models

1. First-Order Model (Screening)

$Y = \beta_0 + \sum_{i=1}^{k} \beta_i X_i + \epsilon$

2. Second-Order Quadratic Model (Optimization)

$Y = \beta_0 + \sum_{i=1}^{k} \beta_i X_i + \sum_{i=1}^{k} \beta_{ii} X_i^2 + \sum_{i<j} \beta_{ij} X_i X_j + \epsilon$

Where $\beta_{ii}$ parameters model non-linear quadratic curvature.

---

Core Response Surface Designs

1. Central Composite Design (CCD)

Constructed by adding axial (star) points and center points to a standard $2^k$ factorial design:

• Factorial Points: $2^k$ points at coded levels $(\pm 1, \pm 1)$.
• Axial Points: $2k$ points located at distance $\alpha$ along factor axes: $(\pm \alpha, 0), (0, \pm \alpha)$.
• Center Points: $n_c$ replicate points at $(0,0)$ to estimate pure experimental error.

Choice of Axial Distance $\alpha$

• Rotatability: A CCD is rotatable when $\alpha = (2^k)^{1/4}$. Rotatability ensures equal variance of predicted values at all points equidistant from design center.

2. Box-Behnken Design (BBD)

An independent spherical second-order design that places experimental runs at midpoints of edges of process space.

• Key Operational Advantage: BBD does NOT contain extreme factor combinations (all points lie within factor boundaries), making it ideal when extreme combinations cause safety trips or equipment damage.

---

Path of Steepest Ascent & Ridge Analysis

When starting far from the optimal region, the Black Belt fits a first-order model and follows the Path of Steepest Ascent to rapidly move toward the optimal region:

$\text{Step Size in Factor } X_i = \Delta X_i \propto \hat{\beta}_i$

Once near stationary point, canonical ridge analysis identifies whether stationary point is a Maximum, Minimum, or Saddle Point using eigenvalues of second-order coefficient matrix $B$.

---

Detailed Canonical Analysis & Stationarity Assessment

When evaluating a second-order response surface model:

$Y = \beta_0 + \sum_{i=1}^{k} \beta_i X_i + \sum_{i=1}^{k} \beta_{ii} X_i^2 + \sum_{i<j} \beta_{ij} X_i X_j$

The stationary point $\mathbf{x}_0$ where the estimated response gradient equals zero is calculated in matrix notation as:

$\mathbf{x}_0 = -\frac{1}{2} \mathbf{B}^{-1} \mathbf{b}$

Where $\mathbf{b}$ is the vector of first-order regression coefficients ($\hat{\beta}_1, \hat{\beta}_2, \dots, \hat{\beta}_k$) and $\mathbf{B}$ is the symmetric matrix of second-order pure quadratic and cross-product interaction coefficients:

$\mathbf{B} = \begin{bmatrix} \hat{\beta}_{11} & \frac{1}{2}\hat{\beta}_{12} & \dots \ \frac{1}{2}\hat{\beta}_{12} & \hat{\beta}_{22} & \dots \ \dots & \dots & \ddots \end{bmatrix}$

Canonical Transformation & Eigenvalue Analysis

To determine the geometric nature of the stationary surface, canonical transformation translates the coordinate origin to $\mathbf{x}_0$ and rotates factor axes along the principal axes of curvature:

$w_i = \lambda_1 w_1^2 + \lambda_2 w_2^2 + \dots + \lambda_k w_k^2$

Where $\lambda_1, \lambda_2, \dots, \lambda_k$ are the eigenvalues of matrix $\mathbf{B}$:

• Maximum Point: All eigenvalues are strictly negative ($\lambda_i < 0$). The surface is a dome; moving in any direction decreases $Y$.
• Minimum Point: All eigenvalues are strictly positive ($\lambda_i > 0$). The surface is a bowl; moving in any direction increases $Y$.
• Saddle Point: Eigenvalues have mixed signs (some positive, some negative). The surface resembles a mountain pass.

---

Comparison of Second-Order Response Surface Designs

---

Practical Multi-Response Optimization: Desirability Functions

In real industrial projects, Black Belts must optimize multiple competing responses simultaneously (e.g., maximizing weld tensile strength $Y_1$ while minimizing weld cycle time $Y_2$ and cost $Y_3$). The Harrington Desirability Function methodology transforms each predicted response $\hat{Y}_i$ into an individual desirability scale $d_i \in [0, 1]$:

Composite Overall Desirability ($D$)

$D = \left( d_1^{w_1} \times d_2^{w_2} \times \dots \times d_m^{w_m} \right)^{\frac{1}{\sum w_i}}$

Where $w_i$ represents the relative importance weight assigned to response $Y_i$. Numerical optimization algorithms search factor space to maximize composite desirability $D$. If any single response fails to meet its critical specification limit ($d_i = 0$), the overall composite desirability drops to $D = 0$.

---

Central Composite Design (CCD) Types: CCC, CCI, and CCF

Depending on process boundaries and factor limits, Black Belts select from three structural CCD variations:

1. Central Composite Circumscribed (CCC): Standard CCD where star points extend beyond the factor bounds ($\alpha > 1.0$). Provides true rotatability across a circular/spherical design region.
2. Central Composite Inscribed (CCI): Factorial limits are scaled inward so star points lie on the actual factor boundaries ($\alpha = 1.0$). Used when factor bounds cannot be exceeded.
3. Central Composite Face-Centered (CCF): Star points are placed at the centers of factor faces ($\alpha = 1.0$). Requires only 3 levels per factor ($-1, 0, +1$), simplifying physical equipment setup.

---

Sequential Experimentation Strategy in DMAIC

DMAIC Experimental Progression:
1. Screening Phase (Fractional Factorial 2^(k-p))  --> Filter vital few factors (k -> 3)
2. Characterization Phase (Full Factorial 2^k)    --> Model main effects & 2-way interactions
3. Optimization Phase (Central Composite / BBD)   --> Model quadratic curvature & find Y_max