At the top of the GeoDict-AI Design of Experiments dialog, the name for the files containing the design results can be entered for Result File Name (*.gdr). The default name can be kept, or a new name can be chosen to fit the current project.

For Generation Script, browse to the *.py file containing the information to generate the training data. How to create this GeoPy file is explained here.

By default the built-in example makeGrainsGeneration.py is loaded. To learn more about this example refer to the Create Training Data topic.

The generation script must contain the parameter gdSeed. It should be used, to provide a random seed to any stochastic structure generation commands (e.g. FiberGeo or GrainGeo). Enter a Start Value for gdSeed. Starting with the entered value, with each structure generation gdSeed is counted one up.

After selecting a generation script, the Generation Script Parameters panel appears below the gdSeed value. If the generation script contains variables other than the gdSeed parameter, here, the allowed value ranges for these parameters can be given in the min and max parameter boxes.

In the example below, the selected Generation Script contains the parameters gdSeed, gd_SVP and gd_BinderSVP. According to the entered check values, which are taken automatically in the Design of Experiments dialog, the solid volume percentage should be chosen between 20% and 80% and the Binder SVP should be chosen between 2% and 15%. In the example, we then decided, that the solid volume percentage will be varied between 30% and 40%, and the binder solid volume percentage will vary between 1.5% and 2.5%.

Select a Variation method for the script parameters to determine how the parameters are chosen from the entered value ranges. The available number generation methods are Sobol and Random.

Random will draw independent random samples in the given parameter ranges. The start value for gdSeed will be used to initialize the random number generation. While all other settings are left the same, the same gdSeed start value always generates the same parameter sequences. Accordingly, different seeds generate different parameter sequences.

Sobol uses a so-called sobol sequence for sampling which generally results in more uniform covering of the n-dimensional parameter space. Sobol does not depend on gdSeed. To learn more about this algorithm refer to Sobol, 1967.

In the following example, observe the differences between the two Variation options Sobol and Random.

The example script had three parameters. The Start Value for gdSeed was set to 5 and the parameter ranges were set to 30-40 for Solid Volume Percentage and 1.5 – 2.5 for Binder SVP.

The Number of Training Structures was set to 100 and the Number of Test Structures were set to 10.

This results in the following plots for the two different Variation options. Observe how the parameter values are selected uniformly in the given ranges, if Sobol is selected, while they are chosen randomly, if Random is selected.

The Training Structures will be used later to train a neural network. For each parameter combination an input and an output file are generated. The neural network will learn how to transform the input file in the given output file. Learn more about input and output files here. Select the Number of Training Structures. The more structures used, the more the neural network can learn in the training. The minimum needed number depends on the number of varying structure parameters, but in most cases 100 training structures lead to good results.

The Test Structures can be used to validate a neural network, that was trained on the training structures. They are created with the same generation script but with different parameter combinations within the given boundaries. If the neural network performs well on these test structures, it should also deliver good results on segmented scans with the corresponding statistical properties. The performance is validated, by applying the trained network to the input files and comparing the results to the corresponding output files. Learn more about input and output files here. Select the Number of Test Structures. The more structures used, the better for the neural network validation. In most cases with 10 test structures the network can be validated well.

However, for testing of the generation script it is recommended to choose a small Number of Training and Test Structures, e.g. 1-5.