Center Points in DOE: Detecting Curvature in Your Process Model

Design of Experiments provides a disciplined way to learn how processes behave. It replaces intuition with evidence. It also reduces the risk of making large changes based on weak assumptions. However, many DOE studies quietly rely on one fragile belief. They assume responses change linearly across factor ranges. That assumption often fails. Chemical reactions saturate. Physical systems hit constraints. Human-driven processes adapt. As a result, response behavior bends. Curvature appears. That’s where center points come in.

When curvature exists, linear DOE models mislead. They distort effect estimates, hide risk, and push optimization toward unsafe extremes. Center points exist to expose this problem early.

This article explains center points and builds from fundamentals to execution. It explains what center points are, why they matter, how they work, how to plan them, and how to interpret results. By the end, you will understand why center points protect both learning and decision-making.

What Curvature Means in DOE

Curvature describes a nonlinear relationship between factors and responses. The response does not change at a constant rate as factor levels increase or decrease. Instead, the slope bends.

In practical terms, curvature means the effect of a factor depends on location. A small change near the low end may have a large effect. The same change near the high end may have little impact. Sometimes the response peaks and then declines.

From a modeling perspective, curvature signals second-order behavior. Quadratic terms matter. The true response surface curves through the design space.

Two-level DOE designs cannot capture this behavior directly. With only low and high levels, the model fits a straight line between extremes. That line may miss the true behavior in the middle.

Curvature matters because it alters conclusions. Ignoring it leads to overconfident predictions, unstable optima, and poor long-term performance.

Why Curvature Matters in Process Learning

Curvature introduces structural error into DOE models. Structural error differs from random noise. Noise adds uncertainty. Structural error biases results.

When curvature exists, linear models misattribute effects. Main effects absorb curvature. Interactions appear inflated. Slopes exaggerate gains near extremes.

These errors influence decisions. Teams may push processes too hard. They may select operating points that perform well briefly but fail over time. They may redesign systems based on misleading trends.

Curvature also undermines extrapolation. Linear models encourage prediction beyond tested ranges. Curved systems punish that behavior.

Center points reduce these risks. They test whether the linear assumption holds before teams commit to conclusions. That safeguard improves both technical accuracy and business outcomes.

The Linear Assumption in Two-Level DOE

Two-level factorial designs dominate early experimentation. They screen many factors efficiently, minimize run count, and remain easy to analyze and explain.

However, two-level designs assume linear behavior by construction. Each factor appears only at a low and high level. The model estimates slopes but cannot bend.

This limitation often goes unnoticed. Teams interpret slopes as universal truths. They assume “more is better” or “less is better” across the entire range.

In reality, many processes behave linearly only over narrow regions. Outside those regions, curvature dominates.

Center points add a reality check. They test whether linearity holds within the chosen factor ranges. They preserve the efficiency of two-level designs while protecting against false confidence.

What Center Points Are

Center points are experimental runs placed at the midpoint of all numeric factors. Each factor uses the average of its low and high values.

If temperature ranges from 50 to 90, the center point uses 70. If pressure ranges from 100 to 300, the center point uses 200. All numeric factors move to their midpoints simultaneously.

Center points represent nominal operating conditions. They often reflect how the process normally runs.

Unlike factorial points, center points do not explore extremes. Instead, they probe the interior of the design space. This interior position allows them to detect curvature efficiently.

Center points add information about shape without adding a full third level to every factor.

What Center Points Are Not

Center points do not create a response surface model. They do not estimate quadratic coefficients; they only test whether quadratic behavior exists.

They also are not simply replicates, even though multiple center points often appear in a design. Replicates reduce uncertainty. Center points challenge model adequacy.

Confusing these roles weakens experiments. Some teams add center points only for error estimation. Others ignore curvature results entirely.

Center points exist to test assumptions. That purpose should guide how teams design, analyze, and interpret experiments.

Conceptual Logic Behind Center Points

The logic behind center points remains intuitive. In a linear system, the response changes at a constant rate. Because of that, the midpoint response equals the average of the responses at the extremes.

If the midpoint response differs consistently, the system bends.

Center points measure that difference directly. They compare reality to the linear prediction.

This logic does not depend on factor count, interactions, or noise structure. It relies on geometry.

Because of this simplicity, center points remain easy to explain to engineers, operators, and leaders alike.

Geometric Interpretation of Curvature

Geometry provides a clear mental model. Picture a straight line connecting the low and high responses. That line passes through the midpoint.

Now imagine the true response surface curves upward or downward. The midpoint of the curve no longer lies on the straight line.

The vertical distance between the straight line and the curve at the midpoint represents curvature.

Center points sample that location directly. They reveal whether the linear approximation misses reality.

This geometric view explains why center points work across many DOE designs. They test shape, not coefficients.

Statistical Basis of the Curvature Test

Statistically, curvature testing compares two averages. One average comes from the factorial points. The other comes from the center points.

Under linear behavior, these averages should match within random error. A systematic difference indicates missing second-order terms.

The analysis partitions variation into:

• Variation explained by linear effects
• Variation attributed to curvature
• Pure error

If curvature explains more variation than pure error, the test becomes significant. That result signals model inadequacy.

This framework keeps the test focused. It avoids overfitting while answering a critical question.

Most statistical DOE software, including Minitab and JMP, reports a formal curvature test when center points are included, allowing teams to assess model adequacy directly.

Center Points and Pure Error Estimation

Multiple center points create repeated measurements at identical settings. These repeats estimate pure error.

Pure error reflects inherent process variability. It excludes model structure.

Pure error strengthens lack-of-fit testing. It helps distinguish between noise and systematic mismatch.

This dual role increases efficiency. A few center points provide both curvature detection and error estimation.

Because of this efficiency, center points often replace some replicates in early DOE stages.

Difference Between Center Points and Replicates

Replicates repeat the same run conditions. They reduce uncertainty. They improve estimate precision.

Center points test model structure. They evaluate whether the assumed shape fits reality.

A design with many replicates but no center points may appear precise while remaining wrong. The estimates look stable but biased.

Both tools matter. However, they solve different problems. Effective DOE strategies use both intentionally.

When Should You Include Center Points?

Center points belong in most DOE studies with numeric factors. Screening experiments benefit first. Early detection prevents false confidence.

Characterization experiments also benefit. They confirm whether linear models remain valid across ranges.

Even confirmation studies gain value. Center points verify assumptions before scale-up.

As a rule, include center points unless physics guarantees linearity across the range.

How Many Center Points Are Enough?

Detection power depends on count. Too few center points weaken the curvature test.

Most practitioners use three to five center points. Highly variable processes may require more.

DOE Size	Center Points	Purpose
Small	3	Basic detection
Medium	4–5	Stable testing
Large	5–7	High confidence

Planning this early prevents underpowered diagnostics.

Center Points in Full Factorial Designs

Full factorial designs estimate all interactions. They still assume linearity.

Center points test that assumption without expanding the design excessively.

Because full factorials often involve fewer factors, curvature detection becomes clearer. The midpoint comparison remains clean and interpretable.

Center Points in Fractional Factorial Designs

Fractional factorial designs trade information for efficiency. Aliasing hides interactions. Curvature hides even deeper.

Center points provide independent shape information. They do not interfere with alias structures.

This independence makes center points especially valuable in screening experiments.

Blocking Considerations for Center Points

Blocking controls known nuisance variation such as batches or shifts.

Center points should follow the same blocking strategy as factorial runs. Ideally, each block includes center points.

This approach prevents confounding curvature with block effects. It preserves interpretability.

Randomization and Center Points

Randomization protects against time-related bias. Center points require the same protection.

Clustering center points invites drift effects. Spreading them across the run order avoids this risk.

Execution discipline matters as much as design structure.

Interpreting a Significant Curvature Test

A significant curvature result means the linear model fails within the tested range.

It does not describe the shape nor locate the optimum.

It simply signals the need for a more flexible model.

Options After Detecting Curvature

Teams have options. One option narrows factor ranges to regain local linearity.

Another option escalates to response surface methods. Quadratic models capture curvature explicitly.

The right choice depends on cost, urgency, and learning objectives.

Interpreting a Non-Significant Curvature Result

A non-significant result supports linear modeling within the tested range.

Confidence increases. Decisions become safer.

However, extrapolation remains risky. Curvature may still exist outside the range.

Conclusion

Center points play a critical but often overlooked role in Design of Experiments. They do not optimize processes or estimate effects. Instead, they protect the assumptions behind every conclusion.

Most DOE studies rely on linear models. Real processes rarely behave that way. Physical limits, chemical saturation, and human adaptation introduce curvature. When that curvature goes undetected, linear models mislead. Decisions drift toward extremes. Performance suffers.

Center points provide an early warning. They test whether the midpoint behaves as a straight-line model predicts. When it does not, they signal the need for a better model while change remains inexpensive.

They also support disciplined experimentation. Center points guide when to escalate to response surface methods. They justify narrowing ranges or expanding model complexity.

Most importantly, center points build confidence. They show that assumptions were tested, not ignored. They make DOE results easier to trust and defend.

Center points may seem minor. Their impact is not.