I mentioned the process automation concept of ISight in a previous simulation automation blog. ISight is an open source code simulation automation and parametric optimization tool to create workflows that automate the repetitive process of model update and job submission with certain objectives associated with it. The objective could be achievement of an optimal design through any of the available techniques in ISight: Design of experiments, optimization, Monte Carlo simulation or Six Sigma. In this blog post, I will be discussing various value added algorithms in DOE technique; I will discuss other techniques in future blogs.
Why design of experiments
Real life engineering models are associated with multiple design variables and with multiple responses. There are two ways to evaluate the effect of change in design variable on response: Vary one at a time (VOAT) approach or Design of experiments (DOE) approach. The VOAT approach is not viable because:
- This approach ignores interactions among design variables, averaged and non-linear effects.
- In models associated with large FE entities, each iteration is very expensive. VOAT does not offer the option of creating high fidelity models with a manageable number of iterations.
With the DOE approach, user can study the design space efficiently, can manage multi dimension design space and can select design points intelligently vs. manual guessing. The objective of any DOE technique is to generate an experimental matrix using formal proven methods. The matrix explores design space and each technique creates a design matrix differently. There are multiple techniques which will be discussed shortly and they are classified into two broad configurations:
- Configuration 1: User defines the number of levels and their values for each design variable. The chosen technique and number of variables determines number of experiments.
- Configuration 2: User defines the number of experiments and design variables range.
This is a three level factorial design consisting of orthogonal blocks that excludes extreme points. Box-Behnken designs are typically used to estimate the coefficients of a second-degree polynomial. The designs either meet, or approximately meet, the criterion of rotatability. Since Box-Behnken designs do not include any extreme (corner) point, these designs are particularly useful in cases where the corner points are either numerically unstable or infeasible. Box-Behnken designs are available only for three to twenty-one factors.
Central Composite Design Technique […]