Design of experiments (DOE) calculator
This calculator plans a design of experiments for a simulation study. Choose what you want to learn, then enter the number of variables, the time per run and how many runs can go at the same time. It gives the design, the number of runs, the wall time and the points to run. Each count is listed with its source under Run-count formulas.
Your study
Work it out from cores
Name the variables and their ranges
The names become the CSV column headers. For a Design Manager import, use the parameter names exactly.
| Name | Min | Max | Unit |
|---|
Advanced
Leave this off for a deterministic solver, which gives the same result for the same inputs.
Your design
Runs
The table and the CSV are step 1. Step 2 depends on which variables the screen keeps. Open it with ‘Use this design’ once you know them.
Comma-separated, one column per variable, headed by the variable name. In Design Manager, import it into a Manual study with Import CSV.
Screen first
The steps run one after the other, and their wall times add up.
Step 1 is a definitive screening design, because a two-level screen can miss a variable whose effect is mostly curved.
You set the budget
Simcenter STAR-CCM+ Design Manager and Simcenter HEEDS spend the number of designs you give them. The calculator shows what your budget allows, not a run count of its own.
Siemens’ rule of thumb for a single objective is about 10 evaluations per variable below 10 variables, 8 per variable from 10 to 20 and 5 above 20. A Pareto search needs a multiple of its archive size, see STAR-CCM+ and HEEDS.
Adaptive sampling in Design Manager · Optimization Intelligence in Simcenter HEEDS · Ask for a demo with your own problem
Points to run
Your design in two variables
Compare and check
Compare the alternatives
| Design | Runs | Wall time | Levels | Estimates | Corner runs | Fits budget |
|---|
Show the calculation with your numbers
What is design of experiments?
Design of experiments (DOE) is a way of planning a set of runs in advance so that the results can be analysed to give valid, objective conclusions. In a simulation study you change the design variables on purpose and watch how the results respond. A well-chosen design gets as much information as possible out of the runs you can afford. That number is usually set by the time each run takes.
How to use the calculator
Start with the goal, because the right design depends on what you want to learn and on how many variables you have.
Under Your design the calculator shows the design, the runs and the wall time, with the reason. With a deadline or a run limit, it also shows whether the design fits. From eight variables it suggests screening first when the goal is interactions or curvature, and from 13 when it is a map.
Give the variables names and ranges to get the points in your own units, as a CSV file, as a table for Excel or as a link that opens the calculator with your inputs.
Why a full factorial does not scale
A full factorial design runs every combination of the levels of all the variables. With k variables at L levels that is Lk runs, and with mixed levels the product, such as 2 × 3 × 5 = 30. Six variables need 64 runs at two levels and 729 at three.
At 2 h per run and four at a time, the 729 runs take 183 batches, about 15 days. Ten variables at three levels need 59 049 runs, about 3.4 years. The designs below are each built for one goal and need far fewer runs (figure 3).
CCD, face-centred
Box–Behnken
DOE designs compared
Two-level factorial and screening designs find which variables matter, and the full factorial and the resolution V fractions also show how they interact. Response surface designs fit curvature with a quadratic model. Latin hypercubes and Sobol sequences spread the runs through the whole space for a surrogate model. Figure 2 shows the main designs for three variables.
Full and fractional factorial designs
A two-level full factorial, 2k runs, estimates every main effect and interaction, and the calculator uses it up to four variables. A fractional factorial design runs a half, a quarter or a smaller fraction of the 2k runs of the full factorial, and sets p of the variables from products of the others, the generators. The shortest word in their defining relation gives the resolution.
The resolution tells you which effects are aliased, or mixed up so that the design cannot tell them apart. In a half fraction of three variables with C = AB, the effect you estimate for C is really C plus the AB interaction.
| Resolution | Main effects | Two-variable interactions | Runs, 6 variables |
|---|---|---|---|
| III | Aliased with two-variable interactions | Aliased with main effects | 8 |
| IV | Clear of two-variable interactions | Aliased with each other | 16 |
| V or higher | Clear | Clear, if three-variable interactions are negligible | 32 (resolution VI) |
With two levels a fractional factorial cannot fit curvature, and a centre run only shows that some curvature is there.
Plackett–Burman and definitive screening designs
Plackett–Burman designs (Plackett and Burman, 1946) come in multiples of four runs and screen up to one variable fewer than their runs. Nine to eleven variables need 12 runs, against 32 at resolution IV. The price is aliasing with two-variable interactions. In the 8- and 16-run designs, which are regular resolution III fractions, the aliasing is complete. In the 12-, 20- and 24-run designs it is partial.
A definitive screening design (Jones and Nachtsheim, 2011) tests each variable at three levels in 2k + 1 runs for an even k, 2k + 3 for an odd k and at least 13. Main effects come out clear of every two-variable interaction and squared term, and from six variables any three of them can be fitted with a full quadratic.
In the calculator, the goal Find which variables matter compares these screening designs for your number of variables, and Estimate interactions gives the smallest resolution V fraction.
Response surface methodology (RSM)
Response surface methodology (RSM) fits a model of the result to a designed set of runs, usually a quadratic polynomial. The fitted surface shows the shape of the response within your ranges. Use it when you expect curvature, such as an optimum inside your ranges. NIST’s selection table gives a central composite or Box–Behnken design for two to four variables and says to screen first from five.
The quadratic model has a constant, a linear and a squared term per variable and a term per pair of variables. That is q = (k + 1)(k + 2)/2 terms, 28 for six variables, and a design needs at least q distinct runs. The calculator shows how many are left over to check the fit.
A squared term needs at least three levels per variable. A two-level factorial has only two, while the central composite and Box–Behnken designs below have three or more. If you have already run a full factorial or a resolution V fraction in the same variables, a central composite design only needs the axial runs and a centre run on top. A resolution IV screen is not enough, because its two-variable interactions are aliased with each other. In the calculator this is the goal Fit curvature.
Central composite design (CCD)
A central composite design (Box and Wilson, 1951) adds two axial runs per variable and a centre run to a two-level factorial core. The core is the full 2k up to four variables and the smallest resolution V fraction above. Six variables take 32 + 12 + 1 = 45 runs.
A rotatable design puts the axial runs at α = F¼, 1.682 for three variables and 2.828 for seven, outside your ranges. Simulation ranges are often real limits, such as a geometry that stops meshing, so we make the face-centred design with α = 1 the default. All runs stay inside the ranges, at the price of less precise squared terms.
Box–Behnken design
A Box–Behnken design (Box and Behnken, 1960) fits a full quadratic model with three levels per variable. Its runs sit at the edge midpoints and the centre. No run sets every variable to an extreme, which helps when the corner designs cannot be meshed. With one centre run it needs 13 runs for three variables against 15 for the CCD. From four to six variables the CCD needs as many or fewer, but at seven a Box–Behnken design needs 57 runs against 79, because the CCD core jumps to 64.
Sobol sequences
A Sobol sequence (Sobol, 1967) is a deterministic, quasi-random sequence that fills the space more evenly than random points. Its balance holds only at a power of two. This means that the calculator rounds 10k up, to 64 for six variables. A Sobol plan grows by continuing the sequence, from 64 to 128 runs, and keeps every run. A Latin hypercube can be doubled and keep its runs only with a special construction (Sallaberry, Helton and Hora, 2008), which the calculator does not use. The goal Map the space for a surrogate gives the Sobol points as CSV.
What is Latin hypercube sampling (LHS)?
Latin hypercube sampling (LHS) is a way to choose N sample points so that, for every variable, each of N equal slices of its range holds exactly one sample. McKay, Beckman and Conover (1979) introduced it as a better alternative to plain random sampling for studies with computer codes.
To build one with N runs:
- Split the range of each variable into N equal slices.
- Shuffle the order of the slices, separately for each variable.
- Pair the shuffled slices row by row, so that each run gets one slice of every variable.
- Take a value inside each slice. The calculator takes the midpoint.
In the example below, the inlet velocity lies between 1 and 5 m/s and the wall temperature between 300 and 400 K. Five runs use every fifth of both ranges once.
| Run | Inlet velocity, m/s | Wall temperature, K |
|---|---|---|
| 1 | 1.0–1.8 | 360–380 |
| 2 | 1.8–2.6 | 300–320 |
| 3 | 2.6–3.4 | 380–400 |
| 4 | 3.4–4.2 | 320–340 |
| 5 | 4.2–5.0 | 340–360 |
Random
Latin hypercube, maximin
Sobol
Powers of two, the sizes at which Sobol points keep their balance.
A grid with five values per variable needs 25 runs to do the same, and it also covers every pair of values. Note that a Latin hypercube only promises the slices of each variable on its own. A diagonal design passes that test and leaves most of the square empty. The calculator therefore optimises the pairing.
A maximin Latin hypercube (Morris and Mitchell, 1995) keeps one run per slice and makes the smallest distance between two runs as large as possible. Simcenter STAR-CCM+ Design Manager also optimises its Latin hypercube for an even spread. Figure 4B shows that distance next to the one for random points. The goal Map the space for a surrogate builds a maximin Latin hypercube for your own variables and ranges.
Random sampling has no rule for the slices. Points can bunch up and leave parts of a range untested. In our comparison of sampling methods in Design Manager, with seven variables and 90 designs, a Latin hypercube gave a more accurate surrogate than random sampling.
About ten runs per variable is a common first size for a Gaussian-process (kriging) surrogate (Loeppky, Sacks and Welch, 2009), and the calculator starts there. A quadratic polynomial needs at least q runs, which is more than 10k from 17 variables (171 against 170). After a screen, count only the variables that matter. Loeppky and his co-authors also found that when ten runs per variable give a poor surrogate, more runs tend to help slowly.
Choose a design
Start from what you want to learn, then check the runs against your cores and your deadline. The runs below are for six variables.
| Goal | Designs | Runs |
|---|---|---|
| Find which variables matter | Fractional factorial, resolution IV / Plackett–Burman / definitive screening | 16 / 8 / 13 |
| Estimate interactions | Fractional factorial, resolution V or higher (here VI) | 32 |
| Fit curvature | Face-centred CCD / Box–Behnken, one centre run | 45 / 49 |
| Map the space for a surrogate | Maximin Latin hypercube / Sobol | 60 / 64 |
| Find the best design | Optimisation or adaptive sampling in Simcenter STAR-CCM+ Design Manager or Simcenter HEEDS | The budget you set |
| Many variables, from about 8 | Screen first, then the design for your goal on the variables that matter | 12 variables, 4 that matter: definitive screening, then a CCD: 25 + 25 = 50 |
A common sequence is to screen first, drop the variables that do not matter and fit or map the rest.
Run-count formulas
The calculator uses these counts for k variables, a time trun per run and c runs at the same time.
N = Lk, or L1 × L2 × … × Lk (1)
Full factorial (NIST/SEMATECH 5.3.3.3 and 5.3.3.10).
N = 2k−p (2)
Fractional factorial, from NIST Table 3.17 up to k = 11. For k = 12–20, resolution IV in 32 or 64 runs (Webb, 1968) and resolution V in 256 or 512 (Xu, 2009), with our own generators, checked by word length.
N = 4⌊k/4⌋ + 4 (3)
Plackett–Burman (NIST/SEMATECH 5.3.3.5).
N = 2m + 1, with m = k for even k, k + 1 for odd k, and at least 6 (4)
Definitive screening, minimum runs (Jones and Nachtsheim, 2011; JMP; Minitab).
q = (k + 1)(k + 2)/2 (5)
Terms of a full quadratic model (Box and Behnken, 1960). With interactions only, 1 + k + k(k − 1)/2, and linear, k + 1.
N = F + 2k + nc (6)
Central composite design for k = 2–10, with F = 2k up to k = 4 and the smallest resolution V fraction above. α = 1 face-centred, F¼ rotatable, √k spherical (NIST/SEMATECH 5.3.3.6.1; Stat-Ease).
N = 12, 24, 40, 48, 56 + nc for k = 3, 4, 5, 6, 7 (7)
Box–Behnken (Box and Behnken, 1960; NIST/SEMATECH Table 3.28; Minitab for seven variables).
N = 10k, or 2⌈log₂ 10k⌉ for Sobol (8)
Latin hypercube for a kriging surrogate (Loeppky, Sacks and Welch, 2009). Sobol at the next power of two (SciPy), with the Joe and Kuo (2008) direction numbers.
twall = ⌈N/c⌉ trun (9)
Wall time in whole batches of c runs.
Nmax = c ⌊T/trun⌋ (10)
The most runs that finish within a deadline T.
Note that we count one centre run in the CCD and the Box–Behnken design. A deterministic simulation gives the same result for the same inputs (Sacks et al., 1989). A repeat adds no information. In a physical test, repeated centre runs measure the scatter. With the noisy-solver box ticked, the calculator takes the centre runs implied by NIST Table 3.28, which are 5, 6, 6, 7 and 10 for a CCD in two to six variables and 3, 3, 6 and 6 for a Box–Behnken design in three to six. Above six variables it takes Minitab’s counts, 10 for a CCD and 6 for a Box–Behnken design. Wall times assume equal run times.
Worked example
This example is made up to show the arithmetic and is not a customer case. A CFD study in Simcenter STAR-CCM+ has six design variables, 2 h per run and eight runs at a time. The aim is a quadratic model, but only a few variables are expected to matter, so we follow NIST and screen first.
Step 1 is the resolution IV fraction 26−2 with E = ABC and F = BCD. The defining relation is I = ABCE = BCDF = ADEF, and the shortest word has four letters. Say three variables matter. Step 2 is a new face-centred CCD in those three, 15 runs for a quadratic with 10 terms.
| Step | Design | Runs | Batches of 8 | Wall time |
|---|---|---|---|---|
| 1. Screen all six variables | 2⁶⁻² resolution IV, E = ABC, F = BCD | 16 | ⌈16/8⌉ = 2 | 2 × 2 h = 4 h |
| 2. Fit the three that matter | Face-centred CCD, 8 + 6 + 1 | 15 | ⌈15/8⌉ = 2 | 2 × 2 h = 4 h |
| Total | 31 | 4 | 8 h | |
| Instead: CCD in all six | Face-centred CCD, 32 + 12 + 1 | 45 | ⌈45/8⌉ = 6 | 6 × 2 h = 12 h |
| Instead: full factorial | 3⁶ | 729 | ⌈729/8⌉ = 92 | 92 × 2 h = 184 h |
Note that step 2 is only as good as the screen. At resolution IV the two-variable interactions are aliased with each other. A variable that acts mainly through an interaction can then look unimportant and be dropped. If you suspect that, keep it. With four variables, step 2 is 25 runs and the whole study 41 runs and 12 h.
A two-level screen can also miss a variable whose effect is mostly curved, such as one with an optimum inside its range. When you expect that, consider a definitive screening design for step 1. It tests each variable at three levels in 13 runs, and the whole study takes 28 runs in the same 8 h. The example keeps the resolution IV fraction to show its generators.
Plan the runs
Runs in parallel
Runs at the same time are your cores divided by the cores per run, or what your licences allow, whichever is lower. A workstation with 32 to 64 cores and 8 to 16 cores per run gives two to eight. Runs go in whole batches, so 17 runs on 16 slots take as long as 32.
Time per run
Use the wall time of a converged run of a similar case, including meshing if the geometry changes with the design variables. For a transient case, count the full simulated time, not one time step.
Check the mesh before the study
Mesh the most extreme designs before you start, such as the corners of a factorial, the axial runs of a CCD with α > 1 and run 1 of a Sobol plan (every variable at its minimum). A CAD Robustness study in Simcenter STAR-CCM+ Design Manager checks that the geometry updates over the parameter ranges without meshing or solving, and it needs no extra licence. A grid study on one design with the GCI calculator shows whether the discretisation error is small next to the differences between designs. Size the first prism layer with the y+ calculator for the design with the highest flow speed.
Design of experiments in Simcenter STAR-CCM+ and Simcenter HEEDS
In both tools a Latin hypercube, adaptive sampling or search study runs the number of designs you set. That is the run count the calculator gives, or the budget you enter for the goal Find the best design.
Design Manager in Simcenter STAR-CCM+
Design Manager comes with Simcenter STAR-CCM+, and Manual and Sweep studies need no extra licence. The DOE study offers two- and three-level full factorials and a Latin hypercube. It needs Intelligent Design Exploration licensing, as do Adaptive Sampling and optimisation (Simcenter STAR-CCM+ documentation, 2606). A Latin hypercube and Adaptive Sampling take their size from the setting Designs to Run.
Without that licence you can run a design from this calculator in a Manual study. Add the input parameters, then choose Import CSV in the study’s Manual Design Table. The columns must match the input parameters, as Florian Vesting’s post Sampling for Design of Experiment in Design Manager (2021) shows. Excel sometimes writes semicolons, so check the file in a text editor.
Since version 2402, Design Manager also has Adaptive Sampling. Without seeds it starts from an optimal Latin hypercube. In our test on a two-variable analytical function, with 150 runs each, it gave a surrogate with 9 % cross-validation error against 26 % for a plain Latin hypercube. An optimisation on either surrogate still missed the true minimum of 0 (31 with adaptive sampling, 107 without), but neither ended in a local minimum.
SHERPA and adaptive sampling in Simcenter HEEDS
Simcenter HEEDS drives the whole loop through its interfaces to CAD and CAE tools, Simcenter STAR-CCM+ among them. Its DOE methods include Latin hypercube and optimal Latin hypercube sampling, partial (fractional) factorial and central composite designs (Simcenter HEEDS documentation). Since HEEDS 2304 it also has a separate Adaptive Sampling study. Its search, SHERPA, runs several global and local methods at once and adapts as it learns.
You set the number of evaluations, and since HEEDS 2510 Optimization Intelligence suggests a minimum. Siemens’ rule of thumb for a single objective, in Simcenter HEEDS and in Design Manager, is about 10 evaluations per variable below 10 variables, 8 per variable from 10 to 20 and 5 per variable above 20, so about 80 for 10 variables and 200 for 40. A Pareto search in Simcenter HEEDS should run at least 15 cycles, which is at least 15 times the archive size, 20 designs by default (Siemens Support Center knowledge base, 2026). For a Pareto study in Design Manager, the Simcenter STAR-CCM+ documentation (2606) suggests 8 to 12 times the archive size, which is 40 by default. The calculator does not estimate a search budget of its own.
A Latin hypercube shows how each variable affects the result, while a search looks for the best design and samples the rest of the space thinly. For a multi-objective search, the Pareto convergence plot in Simcenter HEEDS 2604 shows whether the front has stopped moving. That helps you judge whether more budget is worth it.
If you want help with the first study, ask us for a demo with your own problem. The Design exploration and adjoint optimisation course and Simcenter HEEDS training cover the workflow.
References and changelog
References
NIST/SEMATECH e-Handbook of Statistical Methods, sections 5.1.1, 5.3.3 (Table 3.1), 5.3.3.3–5.3.3.4.7 (Table 3.17), 5.3.3.5, 5.3.3.6–5.3.3.6.3 (Tables 3.21–3.28), 5.3.3.7, 5.3.3.9 and 5.3.3.10. itl.nist.gov
R. L. Plackett and J. P. Burman, “The design of optimum multifactorial experiments”, Biometrika 33(4), 305–325, 1946.
G. E. P. Box and K. B. Wilson, “On the experimental attainment of optimum conditions”, Journal of the Royal Statistical Society B 13(1), 1–38, 1951.
G. E. P. Box and D. W. Behnken, “Some new three level designs for the study of quantitative variables”, Technometrics 2(4), 455–475, 1960.
S. Webb, “Non-orthogonal designs of even resolution”, Technometrics 10(2), 291–299, 1968.
H. Xu, “Algorithmic construction of efficient fractional factorial designs with large run sizes”, Technometrics 51(3), 262–277, 2009, as catalogued in the R package FrF2. CRAN
B. Jones and C. J. Nachtsheim, “A class of three-level designs for definitive screening in the presence of second-order effects”, Journal of Quality Technology 43(1), 1–15, 2011.
M. D. McKay, R. J. Beckman and W. J. Conover, “A comparison of three methods for selecting values of input variables in the analysis of output from a computer code”, Technometrics 21(2), 239–245, 1979.
M. D. Morris and T. J. Mitchell, “Exploratory designs for computational experiments”, Journal of Statistical Planning and Inference 43(3), 381–402, 1995.
C. J. Sallaberry, J. C. Helton and S. C. Hora, “Extension of Latin hypercube samples with correlated variables”, Reliability Engineering & System Safety 93(7), 1047–1059, 2008.
J. Sacks, W. J. Welch, T. J. Mitchell and H. P. Wynn, “Design and analysis of computer experiments”, Statistical Science 4(4), 409–423, 1989.
J. L. Loeppky, J. Sacks and W. J. Welch, “Choosing the sample size of a computer experiment: a practical guide”, Technometrics 51(4), 366–376, 2009.
I. M. Sobol’, “On the distribution of points in a cube and the approximate evaluation of integrals”, USSR Computational Mathematics and Mathematical Physics 7(4), 86–112, 1967.
S. Joe and F. Y. Kuo, “Constructing Sobol sequences with better two-dimensional projections”, SIAM Journal on Scientific Computing 30(5), 2635–2654, 2008.
Software documentation: JMP Help 16.1 on definitive screening designs, Minitab Support on definitive screening, central composite and Box–Behnken designs, Stat-Ease 360 on CCD options, and SciPy on Sobol sequences.
Siemens: the Simcenter HEEDS product page, “What’s new in Simcenter HEEDS 2510” (2025), “What’s new in Simcenter HEEDS 2604?” (2026) and “Unlock the power of CFD simulation and go from 1 to 100s!” (2020), Simcenter blog. Simcenter STAR-CCM+ documentation (2606), Design Manager, and Simcenter HEEDS documentation, in the Siemens Support Center (login required).
F. Vesting, Volupe: “Sampling for Design of Experiment in Design Manager” (2021), “HEEDS 2304 New Adaptive Sampling and MO-SHERPA” (2023), “Efficient design space exploration with adaptive sampling in STAR-CCM+” (2024), “HEEDS Smarter Tools for Better Surrogate Results” (2025), “Optimization Intelligence” (2026) and “Knowing when a Pareto front has stopped moving” (2026).
How we checked the numbers
We ran 556 checks on the calculation engine in Node.js on 28 October 2026. The fractional factorials are checked against NIST Table 3.17 and by the word length of their defining relations. The CCD and Box–Behnken counts are checked against NIST Table 3.28 up to six variables and against Minitab’s summaries above. Every Latin hypercube column is checked for one run per slice, and the Sobol points against a separate implementation. The worked example is also done by hand.
Changelog
October 2026, version 3. The calculator now starts from your goal and gives the design itself, with the points as CSV, a share link and an optional budget. New designs are Plackett–Burman, definitive screening and Sobol, the Latin hypercube is maximin-optimised and the CCD has a choice of α. With one centre run for a deterministic simulation, the CCD for five and six variables drops from 32 and 50 runs to 27 and 45. We removed the illustrative Simcenter HEEDS run count and its formula, because no source supports them.
October 2026, version 2. First version in the CFD Handbook design, with a resolution V core for the CCD from five variables, fractional factorials up to 20 variables and wall time in whole batches.
Next in a CFD set-up
All CFD Handbook tools →GCI = Fs|ε| / (rp − 1)
Calculator
Grid Convergence Index (GCI) calculator
Check that the mesh resolves the differences between your designs before you run the study.
Open calculator →
k–ω SST · k–ε · S–A
Interactive guide
Turbulence model selector
Choose a turbulence model and wall treatment for your Simcenter STAR-CCM+ case, with the checks that go with it.
Open the selector →
y+ = uτy / ν
Calculator
y+ calculator
Size the first prism layer for the design with the highest flow speed in the study.
Open calculator →
Questions about design of experiments
Ask a Volupe specialist
Christoffer Johansson
Senior Application Specialist, M.Sc.
Planning a design study in Simcenter HEEDS or Simcenter STAR-CCM+? We reply within one day.
Ask a technical question →
What is design of experiments (DOE) in simulation?
Design of experiments is a way of planning a set of runs in advance so that the results can be analysed to give valid, objective conclusions. In a simulation study each experiment is one solver run, and you change the design variables on purpose to see how the results respond. A well-chosen design gets the most information out of the runs you can afford.
How many runs does a full factorial design need?
A full factorial runs every combination of levels. With k variables at L levels that is Lᵏ runs, and six variables need 64 runs at two levels and 729 at three. A design built for one goal needs far fewer, for example 16 runs to screen six variables at resolution IV.
How many samples does a Latin hypercube need?
About ten runs per variable is a common first size for a kriging (Gaussian-process) surrogate (Loeppky, Sacks and Welch, 2009), which is 60 for six variables. A quadratic polynomial needs at least (k + 1)(k + 2)/2 runs, which is 28 for six variables. If you have screened, count only the variables that matter.
What is the difference between Latin hypercube sampling and random sampling?
A Latin hypercube with N runs splits each variable’s range into N equal slices and uses each slice exactly once. Plain random (Monte Carlo) sampling has no such rule. Its points can cluster and leave parts of a range untested. In Volupe’s test in Simcenter STAR-CCM+ with seven variables and 90 designs, a Latin hypercube gave a more accurate surrogate than random sampling.
What does the resolution of a fractional factorial design mean?
It tells you which effects are aliased. At resolution III, main effects are aliased with two-variable interactions. At resolution IV, main effects are clear, but two-variable interactions are aliased with each other. At resolution V, main effects and two-variable interactions can all be estimated, assuming three-variable interactions are negligible.
How many runs does a central composite design need?
A central composite design has a two-level factorial core, two axial runs per variable and centre runs. With one centre run, which is enough for a deterministic simulation, that is 9, 15, 25, 27 and 45 runs for two to six variables. A Box–Behnken design needs 13, 25, 41 and 49 for three to six variables, and 57 for seven, where the CCD needs 79.
How many centre runs does a simulation study need?
One, if the solver is deterministic. Repeating a run with the same inputs gives the same result. A repeat adds no information. Repeated centre runs are for physical tests or noisy output, where they measure the scatter. The usual rules give 3 to 6, and NIST’s table implies 10 for a CCD in six variables. A Box–Behnken design needs at least one centre run to fit the model.
How many runs does Simcenter HEEDS need?
You set the budget, the number of evaluations, and SHERPA adapts its search to it. Since HEEDS 2510, Optimization Intelligence suggests a minimum for your problem. Siemens’ rule of thumb for a single objective is about 10 evaluations per variable below 10 variables, 8 per variable from 10 to 20 and 5 above 20, so about 80 for 10 variables and 200 for 40. A Pareto search needs at least 15 times its archive size, which is 20 designs by default in Simcenter HEEDS.
Can I import a design from a CSV file into Simcenter STAR-CCM+?
Yes, into a Manual study, which needs no extra licence. Add the input parameters to the study, then choose Import CSV in its Manual Design Table. The columns must match the input parameters. Name the variables in the calculator after your parameters, because its CSV uses them as the column headers. Volupe’s post on sampling in Design Manager (2021) shows the steps.
Talk to an engineer in your region
Every Volupe business developer holds an engineering degree. Find your regional contact below, or send us a message.
Ask us about DOE calculator
Use this form for licence quotes, demos, training and technical questions.
What happens next
- A reply within one day.
- A scoping call with one of our engineers.
- A quote, demo or other next step that fits your enquiry.
After purchase: support from Volupe simulation engineers.