DOE Design Tool and DOE-POST (Design of Experiments)
GT-SUITE includes comprehensive tools for Design of Experiments (DOE) creation, analysis, and optimization. Users can create a DOE matrix to perform controlled variation of any model inputs (factors) for the purpose of analyzing and predicting model outputs (responses). Full Factorial and Latin Hypercube DOE types are available, or a custom matrix of experiments can be copied from an external source.
The time required to run large DOEs is significantly decreased by utilizing Distributed Computing, which allows models with multiple cases to be divided among many cores of a cluster with multiple execution nodes.
Reviewing Results in Interactive Environment
After running the DOE simulations, the data is read in the DOE Analysis and Optimization Tool in GT-POST. With a guided workflow, the tool takes the user through multiple different views or environments, each dedicated to accomplishing specific tasks in visual and interactive ways. As a platform designed for graphical and intuitive handling of large data sets, the tool conveniently organizes data from multiple operating points that were defined in Case Setup, such as steady-state engine speeds, and a range of flow rates or temperatures. The user can easily review and sort the data, and unwanted data points can be excluded from further analysis.
Automatic data analysis extracts useful information from the data, such as the most influential factors on each response, unimportant factors, and correlations between any two variables.
A central focus of the tool is to train the DOE data to create metamodels, or response surface fits, where the factors serve as input variables, and the responses as the predicted outputs. Available fitting algorithms include polynomials, neural networks, nonlinear interpolation, and data-driven Gaussian processes. Metamodel training is followed by rigorous cross-validation calculations to provide a measure of how well each metamodel generalizes to new data and predicts unknown points. Visual and interactive plots and tables allow review of metamodel quality metrics and side-by-side comparison to determine the best fitting algorithm for each response.
Metamodels provide the ability to make predictions, analyze factor effects on responses, and explore potential designs. XY scatter, 2D contour, and 3D surface plots aid design exploration, where the user receives immediate visual feedback on response predictions as factors are changed. Finally, any one or several metamodels can be directly minimized, maximized, or driven to target values by configuring optimization runs. Multiple different optimizations can be created, configured, run, and archived for a given DOE project. The full collection of capabilities available via the GT-SUITE Integrated Design Optimizer can be utilized.