QE Capability Enhancement (18) | Response Surface Methodology (RSM): Finding the Optimal Parameter Combination

1. A Question That Stumped Me: "The Optimal Solution"

A certain automotive parts company was optimizing parameters to reduce the porosity rate in die-cast components. The QE used a central composite design (CCD) for 19 trials, and after fitting a second-order model, the R² reached 0.94. The software provided the "optimal solution" as: mold temperature 268°C, injection speed 4.2 m/s, and boost pressure 82 MPa, with a predicted porosity rate of 0.8%. On the day the report was submitted, the process engineer asked just one question: "If the mold temperature is increased by 5°C, will the porosity rate continue to decrease or rebound according to your model?" Upon examining the surface plot, it was discovered that this point was rising in the mold temperature direction—it was not a valley but a saddle point: a minimum along one direction and a maximum along another. Setting parameters based on this point would amplify any normal fluctuations into defects, and the prediction interval was wide, ranging from 0.6% to 1.4%. The issue was not with the software but with the lack of verification of the stationary point's nature and the width of the prediction interval, leading to the misinterpretation of a statistical stationary point as the optimal process parameter.

2. Principle: RSM Seeks Stationary Points on the Surface, Not the Best Group in the Mean Values of Trials

The basic assumption of Response Surface Methodology (RSM) is that, given 2 to 4 key factors have been identified, there exists a quadratic relationship between the response and the factors. The model is of the form:

y = b₀ + Σbᵢxᵢ + Σbᵢᵢxᵢ² + Σbᵢⱼxᵢxⱼ

where x is the coded variable (low level -1, high level +1). The optimal point is derived by finding the stationary point of this surface: taking the partial derivatives of each xᵢ and setting them to zero, the solution is the stationary point. Note three important points that determine whether the conclusion is usable.

First, the nature of the stationary point is determined by the signs of the eigenvalues of the quadratic term coefficient matrix. All negative eigenvalues indicate a maximum point, all positive eigenvalues indicate a minimum point, and a mix of positive and negative eigenvalues indicate a saddle point. A saddle point is not an "optimal parameter"; it is merely a mathematical stationary point. It must be clearly labeled in the report and either converted to a ridge analysis (the optimal path along a fixed direction) or retreated to a limited area to find a conditional extremum.

Second, whether the stationary point falls within the experimental region determines the strength of the conclusion. If it falls within the region defined by the cube and star points (an internal solution), it can be directly recommended. If it falls outside the region (an external solution), it can only be traced along the steepest ascent path to the engineering feasible boundary, and the point on the boundary should be selected, with the note that "this is the optimal under constraints, not the surface optimal."

Third, the second-order model is only reliable within the experimental interval. Any predictions outside this interval are extrapolations. The narrower the interval, the less visible the curvature; the wider the interval, the higher-order terms may be incorrectly fitted as second-order terms. Whether the model is sufficient can be determined by three criteria: the lack-of-fit test is not significant (p > 0.05), the difference between the adjusted R² and the predicted R² (PRESS/Q²) is no more than 0.2, and the degrees of freedom for pure error are sufficient.

The final rule: coefficients are only comparable when factors are in coded units; the size of coefficients in original units has no meaning.

3. Five Practical Steps: Each with Criteria

Step 1: Confirm "RSM is Appropriate," Not Just "RSM is Desired." The prerequisite is that 2 to 4 key factors have been identified, the level spacing is ≥ 3 times the equipment resolution, and the response is indeed curved. Criteria: the difference between the mean values of the center points and the corner points is significant (p < 0.05), or preliminary trials suggest that increasing a parameter improves the response, but further increases worsen it; the power for the target effect size (generally 1.5 to 2 times the process standard deviation) is ≥ 0.8. When still searching for the needle in the haystack among 8 factors, start with a fractional factorial screening, not directly with RSM.

Step 2: Select the Design Based on Criteria, Either CCD or Box-Behnken. Both are second-order designs, but they differ in the position and cost of the experimental points:

The repetition of center points is the soul of these designs. Criteria: at least 3 center points, aiming for 4 to 6; degrees of freedom for pure error = number of center points - 1, must be ≥ 3. Fewer than 3 center points render the lack-of-fit test meaningless.

Step 3: Execute the Experiments and Lock Down Data Quality. Randomize the sequence, and use batch blocking if necessary; measure 3 samples under each experimental condition and take the mean to reduce within-group noise; ensure the measurement system is qualified before starting—GR&R P/TV ≤ 10% is required for optimization, 10% to 30% can only show major trends, and greater than 30% requires instrument repair; record the raw material batch, environmental temperature, and humidity for each data row.

Step 4: Fit and Diagnose the Model, Decide Whether It Can Be Used Based on Criteria. Four acceptance criteria: lack-of-fit test p > 0.05 (indicating the second-order shape is sufficient); the difference between adjusted R² and predicted R² is ≤ 0.2 (indicating no overfitting); residuals show no funneling or curvature trends; insignificant terms must be removed one at a time, and the model must be refitted after each removal; do not use old model coefficients to explain new data.

Step 5: Find the Stationary Point, Determine Its Nature, and Conduct Confirmation. After calculating the stationary point, first check the signs of the eigenvalues. If it is a saddle point, convert to ridge analysis or find a conditional extremum within a limited area, do not force the report. If there are multiple responses (such as shrinkage and warpage), use a desirability function or overlapping contour lines to find a compromise area, do not optimize each response separately. Finally, conduct confirmation trials according to the selected combination: independent batches ≥ 3 (for key characteristics, ≥ 5), and the measured mean must fall within the 95% prediction interval of the model; if production capability is also required, the Cpk at this point must be ≥ 1.33 to be included in the control plan and process card.

4. Five Common Misconceptions

Misconception 1: Directly Taking the Software's "Optimal Point" as the Process Optimal. The most common mistake is reporting a saddle point or an external solution as an internal minimum. Criteria: any optimal solution must be accompanied by an explanation of the stationary point's nature and whether it falls within the experimental region.

Misconception 2: Optimizing Only One Response in Multiple Responses. Optimizing a single response often degrades another metric. Reporting only the optimization goal and hiding the sacrificed metric will inevitably lead to repeated issues in mass production. In multiple response scenarios, a compromise area should be provided, and all key metrics should be validated in the confirmation trials.

Misconception 3: Judging Factor Importance Based on Coefficient Size in Original Units. Coefficients are comparable in coded units but not in original units. Changing the unit of the same data can reverse the conclusion.

Misconception 4: Too Few Center Points or Mixing Center Points with Star Points in the Same Batch. Insufficient degrees of freedom for pure error make the lack-of-fit test meaningless; batch effects mixing in can cause the quadratic terms to absorb drift, fitting a non-curved process as curved.

Misconception 5: Focusing Only on R² and Not on Predictive Ability, Self-Validation in Confirmation Trials. An R² of 0.95 may come from overfitting; look at Q²/PRESS. If the confirmation trials use the same batch of materials, the same time period, or even the same data used for modeling, it is self-validation. Independence and the number of batches are strict requirements.

5. Self-Check List

• Key factors are locked down to 2 to 4, level spacing is ≥ 3 times the equipment resolution, and there is evidence of curvature (significant difference between center points and corner points)
• Design selection is based on criteria: use CCD if factors can extend beyond the region and calculate α to ensure rotatability; use BBD if factors cannot exceed the boundary; if experiments are conducted in batches, use orthogonal blocking and place center points in each batch
• Center points ≥ 3 (aim for 4 to 6), degrees of freedom for pure error ≥ 3; sequence randomized, 3 samples measured per condition and mean taken, GR&R P/TV ≤ 10%
• Model acceptance criteria are complete: lack-of-fit p > 0.05, difference between adjusted R² and predicted R² ≤ 0.2, residuals show no anomalies, and the model has been refitted after term removal
• Stationary point nature (maximum, minimum, or saddle point) and whether it falls within the experimental region have been determined; ≥ 3 independent confirmation trials completed, mean values fall within the 95% prediction interval

First, determine the nature of the stationary point, then discuss the optimal parameters.

Knowledge code: 6.4.1

Version: v20260928

Author: QTank QTank is dedicated to providing systematic professional knowledge, methodologies, and practical tools for quality management practitioners, helping enterprises continuously improve their quality capabilities.