Sensitivity Analysis-Integrated Design of Experiments

For a quick, model-free screening of many factors, Morris OAT (elementary effects) is computationally cheap and robust. For a more rigorous variance decomposition that separates main effects from interactions, Sobol indices are preferred but require more runs. In practice, Morris screening is common as a first-pass filter, with Sobol analysis reserved for the refined factor set or the fitted surrogate.

Morris screening needs r × (k + 1) runs, where r is the number of trajectories (typically 10–20) and k is the number of factors. Sobol-based methods need N × (k + 2) runs, where N is a base sample size (often 500–2000 for reasonable convergence). The screening investment is justified when it allows a substantially smaller main DoE.

Yes, but the upfront screening cost is higher for physical runs than for simulations. In practice, teams often use a low-fidelity computational model for screening and reserve physical experiments for the main DoE phase on the reduced factor set. If physical screening runs are feasible, a Plackett-Burman or Resolution III fractional factorial can serve a similar filtering role at lower run counts than a full SA.

Contradictions usually signal that the initial factor ranges were mis-specified, that the screening sample was too small to converge, or that the response is strongly nonlinear in ways the screening design did not capture. Revisit the factor ranges, increase the screening sample, and consider a non-parametric sensitivity estimator before concluding that a factor is truly inert.

No. Plackett-Burman and similar resolution III designs are DoE-only screening tools that estimate main effects by aliasing interactions. SA-DoE is a two-phase framework: the first phase uses sensitivity indices (which are not based on a linear model) to rank factors, and the second phase applies a chosen DoE to the retained factors. SA-DoE is more flexible and can handle nonlinear responses that would confound Plackett-Burman estimates.