21.3.7.5. DOE Selection for Meta-Model

When one selects the DOE method for meta-model construction, Efficient Method is recommended.

../_images/image119.png — Figure 21.95 **DOW & Meta Modeling Methods** dialog box

Guide 1: If the current design is within 50 % range from the center position, ISCD-1 or ISCD-2 is recommended. Otherwise, Discrete Latin Hypercube Design is recommended. The latter is one of stratified random sampling methods. Figure 21.96 shows the recommend range for ISCDs.

../_images/image1201.png — Figure 21.96 The current design location for **ISCDs**

Guide 2: When one uses Discrete Latin Hypercube Design, the number of samples is recommended as more than N=Max {20, 2*k+k*(k-1)/2}, where k is the total numbers of design variables and constants. Also, If N is odd number, then N=N+1 is recommended.

Guide 3: Among ISCDs, ISCD-2 is recommended for Design Optimization. ISCD-1 is done for DFSS/Robust Design Optimization. If a design problem is noisy or complicated, however, then Discrete Latin Hypercube design is recommended.

Guide 4: When the design problem has ‘random constant’ in the DFSS/Robust Design problem, Discrete Latin Hypercube design is recommended. Consider the following examples. It has 2 design variables and 3 random constants. Thus, 2*5+5*4/2=20. The recommended number of sampling points is more than 20.

Guide 5: If the above guidelines fail to converge, then the hybrid sampling method is recommended. First, check Get from simulation history and import all analysis results and check select doe method also, Then, add more sampling points in discrete Latin hypercube design. If you cannot increase the initial sampling points, then another tip is to switch the meta-model method from Kriging to RBF with quadratics or reversely. For this selection, only ‘Get from simulation history’ is recommended.

Guide 6: Although these guidelines can help user’s selection, it cannot guarantee the convergence for all design problems. Consider the mathematical optimization algorithms. Their convergences are sensitive to the initial design values. Similarly, the convergence of meta-model base SAO can be sensitive to the initial sampling points.