The presented experimental designs integrated into the simulation environment have different properties and the optimal choice depends on the type of analysis which has to be performed.
One group of designs is frequently used for fitting a response surface to simulated data. This is called Response Surface Methodology (RSM) and is described in Section 3.2.
The Central Composite design are used to determine the coefficients of a second order responses surface model. Their advantage is the small number of required experiments and for CCI and CCC the property of being a rotatable design. They are preferred when the model evaluations are expensive. It should be noticed when adding new input parameters to a Central Composite design the axial points can not be reused.
Starting from Orthogonal Main Effect the Supplemantary design is aiming to minimize the number of additional experiments required when it is necessary to add new control parameters. This method causes a slightly worse design than a Central Composite Circumscribed design.
Plackett-Burman designs require a low number of experiments, but being a two-level design only linear response surfaces can be extracted.
For fitting response surfaces of higher than second order Three Level Factorial, Latin Hypercube, Random etc. designs can be used to generate the experiments.
To explore the model behavior in a large input parameter space Full Grid or Random designs are used. Here plenty of evaluation points have to be simulated. This implies that the required calculation time for a single evaluation is not to high and the number of available CPUs is sufficient.
For studying the influence of process parameter variations during manufacturing, e.g., for yield analysis, Gauß Random designs are preferable. This type of analysis also requires a large amount of experimental points to receive realistic results.
Design of Experiments methods are also available in other TCAD frameworks namely the Virtual Wafer Fab [47], the TMA Work Bench [59], and the NORMAN framework [5]. Unlike the presented tools these implementations do not use transformations to model the system dependencies.