Service overview
Statistical analysis, designed experiments, regression, reliability and measurement analytics. This service is part of that division and is scoped individually around what you need to learn, decide or publish.
Response surface methodology is the natural next step after screening experiments have shown which factors matter. Instead of asking which factors have an effect, it asks where in the factor space the response is best and how quickly it degrades away from that point. The answer is a fitted surface, usually quadratic, that can be plotted as contours and interrogated numerically.
The method is sequential by design. An initial first-order experiment points in a promising direction, a second-order design maps the neighbourhood of the optimum, and confirmation runs test whether the model predicts reality. It is widely used in process engineering, chemical and environmental studies and manufacturing, where a handful of controllable variables determine yield, strength or cost and where each run is expensive.
If you plan to publish the outcome, see manuscript & publication support, for example computational & simulation manuscripts or original research article writing.
What this service covers
Problems this service can address
- Optimal settings that sit in an unexplored region of the factor space
- Curvature in the response that linear models cannot describe
- Competing responses that need a compromise
Methods and analytical approaches
- Central composite and Box-Behnken designs
- Second-order polynomial model fitting
- Lack-of-fit and diagnostic analysis
- Contour and surface visualisation; desirability-style compromise
Applications
- Process optimisation in manufacturing and chemical systems
- Adsorption, extraction and formulation studies
- Parameter tuning of fabrication processes
Relevant engineering disciplines
Scope, inputs and deliverables
Scope of work
- Review of your objectives, constraints and the evidence already available
- Selection of methods suited to the question and the data
- Analysis or synthesis with documented, repeatable steps
- Written outputs you can reuse in a thesis, proposal or report
- Feasibility check on: factors and feasible ranges
Typical inputs required from you
- Factors and feasible ranges
- Measured responses and replicates
- Preferred direction for each response
Deliverables
- Design plan and fitted models
- Response surface and contour figures
- Predicted optimum with confirmation-run guidance
Highly specialised requests are first checked for feasibility from this information. Please do not send confidential data through the public form; describe it first and agree secure transfer afterwards.
Project workflow and method
- 01Enquiry and requirement assessment
- 02Technical scoping and proposal
- 03Execution
- 04Quality review and delivery, with an agreed round of revisions
Read the full project workflow.
In depth: how the work is done
Choosing the design: central composite and Box-Behnken
A central composite design augments a factorial or fractional factorial core with axial points and centre points so that a full quadratic model can be estimated. Variants differ in whether the axial points lie outside the factorial cube, which extends the explored region but requires factor settings beyond the nominal range. Box-Behnken designs place points at the midpoints of the edges of the cube and avoid extreme corners, which is useful when extreme combinations are unsafe or infeasible. We select the design according to factor count, region of interest, run budget and whether blocking is needed.
Fitting and checking the quadratic model
The second-order model contains linear, squared and interaction terms. After fitting, analysis of variance indicates which terms contribute, and lack-of-fit tests compare the model error with pure error from replicated centre points. A significant lack of fit means the quadratic form is inadequate in the studied region, and the response may need a transformation, a smaller region or a different model. Residual plots reveal non-constant variance and outliers, and coded variables are used to keep coefficients comparable and the numerical problem well conditioned.
Finding and interpreting the optimum
Contour and surface plots show how the response changes in pairs of factors with others held constant. Canonical analysis classifies the stationary point as a maximum, minimum or saddle and reveals ridges along which several settings perform similarly. Such ridges are valuable because they let you trade off cost or robustness without losing performance. When several responses must be satisfied simultaneously, desirability functions or overlaid contour plots identify settings that give an acceptable compromise.
Confirmation and robustness
Predicted optima are hypotheses. Confirmation runs at the proposed setting, ideally with fresh material and a different operator or day, test whether the model's prediction falls within its interval. We also examine how sensitive the response is to small deviations in the factors, since real processes cannot hold settings exactly. A flat optimum is preferable to a sharp one, and the report describes the operating window within which performance remains acceptable.
Planning the sequence of experiments
RSM is rarely a single experiment. A typical sequence begins with a screening design, moves to a steepest-ascent or descent path to reach the right region, and only then fits a second-order model. We help decide how far to move along the path between experiments, when to stop, and how to treat constraints such as maximum temperature or minimum purity. Planning the sequence in advance prevents both wasted runs and the temptation to fit complicated surfaces to data gathered in the wrong place.
Quality control and good practice
Quality-control approach
- Search, selection or analysis steps are documented so another researcher can repeat them.
- Assumptions, exclusions and limitations are stated rather than implied.
- Numbers in text, tables and figures are cross-checked for consistency.
- Conclusions are limited to what the evidence supports.
Common pitfalls this service helps avoid
- Fitting a quadratic far from the optimum. A second-order model is a local approximation; using it over a large region that includes strong higher-order behaviour gives misleading optima.
- No replicated centre points. Without replicates there is no estimate of pure error, so lack of fit cannot be tested.
- Extrapolating outside the design region. Predicted optima lying beyond the experimental region are not supported by data and should be tested before being trusted.
- Optimising one response in isolation. Setting that maximise yield may violate quality or cost limits that matter for the application.
What good reporting looks like
An RSM report states the objective and responses, the factors with coded and natural units, the chosen design and why, the run table and randomisation, the fitted model with ANOVA and lack-of-fit results, residual diagnostics, contour or surface plots, the optimum with prediction intervals and the confirmation results. It also specifies the region within which the model is valid. Conclusions are limited to the tested factors and ranges, and the report states how many runs were replicated and how nuisance variation was handled.
An illustrative scenario
Hypothetical example written for explanation. It is not a client case, a result or a testimonial.
Imagine a team optimising an extraction process with three factors: temperature, time and solvent ratio. Screening showed all three matter. A Box-Behnken design with replicated centre points is run in random order across two days, with day treated as a block. The quadratic model fits without significant lack of fit, and the contour plots show a ridge where several combinations of time and temperature give similar yield. The team picks a setting near the middle of the ridge for robustness and runs three confirmation batches, all of which fall inside the prediction interval.
Questions to settle before you begin
- Have screening experiments already identified the important factors?
- Is the region around the expected optimum safe to explore?
- Can I replicate centre points to estimate pure error?
- Which responses must be satisfied together?
- How will I confirm the predicted optimum with independent runs?
Key terms
- Central composite design
- A design combining factorial, axial and centre points to estimate a quadratic model.
- Box-Behnken design
- A three-level design using edge midpoints that avoids extreme corner runs.
- Lack of fit
- A test of whether model error exceeds pure experimental error.
- Desirability function
- A way to combine several responses into one score for compromise optimisation.
- Coded variable
- A factor rescaled to a standard range, such as minus one to plus one.
Working together on Engineering Data Analysis & Statistics projects
Data analysis engagements start with a data description before any modelling: what was measured, how, under what conditions, with which instruments and with what known problems. Sharing a small sample or a data dictionary first is usually enough to scope the work. Analyses are documented step by step so that every figure and table can be reproduced from the supplied data, and code or workflow files can be provided on request. Please avoid sending confidential or personal data through the public form; secure transfer can be arranged after scoping. Conclusions are phrased with the uncertainty the data justify, not stronger.
Frequently asked questions
Is the predicted optimum guaranteed to hold in practice?
No. It is a model-based estimate within the studied range. We recommend confirmation experiments before reporting it as an optimum.
When should I use RSM instead of a factorial design?
Use factorial or screening designs to find which factors matter. Use RSM once you know the key factors and want to model curvature and locate an optimum.
How many factors can RSM handle?
Practical designs usually handle two to five factors. With more, screening should first reduce the list because the number of runs grows rapidly.
Can RSM be used with simulation results?
Yes. Computer experiments can be used to fit response surfaces, although noise is absent and other designs, such as space-filling plans, may be better suited.
What if the quadratic model does not fit?
Options include transforming the response, narrowing the region, adding terms or using a more flexible model. The cause of lack of fit should be investigated rather than ignored.
How is this service scoped and priced?
After your enquiry we review what you need, ask for missing information and send a written scope and quotation. Price depends on complexity, deliverables and timeline. See the project workflow.