Must the model always be quadratic?

No; the appropriate form depends on the region and evidence.

Does adding more terms always improve decisions?

No; unnecessary flexibility can overfit limited runs.

Can two factors interact?

Yes; the effect of one can vary with the other, which an interaction term represents.

AI Glossary / Statistics and probability / Intermediate

Response surface model

Definition

A response surface model is a mathematical approximation describing how a response changes with two or more input factors over a specified region. Common models include linear effects, interactions, and quadratic terms. In response surface methodology, such models are often fitted to designed experiments to explore operating conditions and identify settings predicted to improve a response.

What it means

The surface is the predicted response above combinations of factor settings. With two factors it can be pictured as a landscape: height represents the response, while horizontal coordinates represent the inputs. With more factors, the same idea extends beyond a simple three-dimensional picture.
The model is usually an empirical approximation over a local region. It need not be the exact physical law of the process, and a predicted optimum is a candidate requiring confirmation.

How it works

Define the response, controllable factors, feasible operating region, and objective. Maximizing yield, minimizing error, and meeting several quality constraints are different tasks. Establish measurement procedures and relevant noise factors before collecting runs.
Choose a design that can estimate the intended terms. A first-order model represents local slopes. An interaction allows the effect of one factor to depend on another. Pure quadratic terms allow curvature along individual factor directions.
A full quadratic model for two factors contains an intercept, two linear terms, one interaction, and two squared terms. The design must support independent estimation of those coefficients. Merely adding center points to a two-level factorial can detect overall curvature without separately identifying all pure quadratic effects.
Code factors when useful, such as mapping a practical low and high setting to -1 and +1. Coding clarifies interpretation and numerical scale, but the relationship to original units must be retained for implementation.
Fit the chosen model and assess residuals, uncertainty, and lack of fit where replication permits. A nonsignificant lack-of-fit test does not prove the model is correct; the design may have limited ability to reveal deviations.
Explore the fitted surface with plots and calculations. A stationary point, where all first derivatives vanish, may be a maximum, minimum, or saddle. Its location must also be feasible and supported by the experimental region.
When the unconstrained optimum is outside the studied region, evaluate boundaries or plan further experiments rather than blindly extrapolating. Multiple responses may require explicit tradeoffs or constrained optimization.
Perform confirmation runs at promising settings, ideally under realistic operating variation. Compare observed results with the model's predictions and update the model if needed. The optimization of a fitted equation is not a substitute for validating the process at the proposed setting.

Three practical examples

AI systems: engineers model the joint effect of two controlled configuration settings on a measurable performance response.
Manufacturing: a team explores how temperature and processing time jointly affect product quality.
Service operations: researchers use a designed trial to approximate how staffing and workload settings influence turnaround time.

Think of it this way

Think of a local map of hills and valleys. The map suggests a promising direction and a possible summit, but it covers only the surveyed territory. Reaching the predicted summit and checking its actual height is the role of a confirmation experiment.

Worked example

Consider an invented quadratic response surface in coded factors x and z:
predicted y = 80 + 2x + 2z minus 2x squared minus 2z squared + xz.
Suppose the studied region is -1 ≤ x ≤ 1 and -1 ≤ z ≤ 1. Setting the partial derivatives to zero gives 2 minus 4x + z = 0 and 2 minus 4z + x = 0. Solving yields x = z = 2/3, which is inside the region.
The curvature matrix has diagonal entries -4 and off-diagonal entries 1. Its eigenvalues are -3 and -5, both negative, so this quadratic has a maximum at that point. Substitution gives predicted y = 244/3, approximately 81.3333. At the design center x = z = 0, predicted y is 80.
Thus the model predicts an improvement of 1.3333 response units over the center. The coefficients are stipulated, not estimated from actual runs, so no uncertainty interval or empirical improvement is implied. A real study would need confirmation and an assessment of whether that improvement matters.

Illustrative case study

Fictional case: Linden Process Lab fits a response surface from a planned experiment and identifies an interior candidate setting. The team checks that the setting is feasible, runs fresh confirmations, and measures variation across operating days. It treats the fitted maximum as a proposal to test rather than immediately replacing the established production setting.

Common misconceptions and mistakes

Do not assume every stationary point is a maximum. A saddle can have zero gradient while improving in one direction and worsening in another.
Do not infer a causal surface from an observational fit without considering confounding. Designed and randomized factor changes offer a stronger basis when properly executed.
A quadratic term can describe local curvature without being valid far outside the data. Avoid optimizing beyond meaningful factor limits.
Response surface methodology is not the same named framework as Shore's response modeling methodology, despite the similar wording.

Business use

Response surfaces help organizations make efficient, structured improvements when experiments are feasible and factors can be controlled. Their value comes from combining design, modeling, uncertainty assessment, and confirmation. Reports should connect predicted improvements to cost, constraints, and reproducible operating settings.

Check your understanding

Why is the worked stationary point an interior candidate?

Answer: Both coded coordinates equal 2/3 and lie between -1 and 1.

Why does the predicted improvement not establish a real gain?

Answer: The example uses a stipulated equation, and a fitted equation would still require validation.

Related terms

Design of experiments
Central composite design
Box - Behnken design
response surface methodology
quadratic model
interaction
stationary point
confirmation experiment

Learn next

Learn next how design geometry determines estimable curvature, how canonical analysis classifies a quadratic surface, and how confirmation experiments test proposed operating improvements.

Sources

Sources: NIST/SEMATECH, Response surface designs, https://www.itl.nist.gov/div898/handbook/pri/section3/pri336.htm ; NIST/SEMATECH, Optimization when there is adequate quadratic fit, https://itl.nist.gov/div898/handbook/pri/section5/pri5314.htm . These primary references describe polynomial response models, design requirements, optimization, and confirmation. The quadratic and fictional case are original.

Previous: Response set Next: Responsible AI