Generating and Applying Surrogate Models

A study runs significantly faster when you compute responses using surrogate models instead of simulations. In Design Manager, a typical application for surrogate models is a Robustness and Reliability study with a large number of sample points in the design space.

Before applying surrogate models, you must first generate suitable surrogates within a design study which covers the part of the design space you are interest in, which can be a local space around the best design or globally scattered samples.

Once the surrogates are generated, they can also be exported to other tools to replace expensive simulations. For essential concepts of surrogates, refer to Surrogate Models.

Surrogate computation and cross validation require an innovatesuite license. See also: Design Manager Licensing.

In Design Manager, surrogates can be computed from any study type except CAD Robustness study type. For most applications, you are advised to use Adaptive Sampling and Design of Experiments (DOE) with Latin Hypercube Sampling method for surrogate generation.

Adaptive Sampling automatically generates the best fit surrogate type according to the Adaptive Strategy you set. For a DOE study, you specify surrogate settings manually.

To generate surrogates with an Adaptive Sampling study:
1. Create a design study.
2. Select the [design study] node and set Study Type to Adaptive Sampling.
3. Select the Adaptive Sampling Settings node and set adaptive sampling properties.
  Refer to Adaptive Sampling for details.
4. Select the Input Parameters node and add input parameters to Parameters.
5. Select the Responses node and add design responses.
  A surrogate uses all the input parameters specified in the design study and provides approximation only for one response. The number of responses you add here determines the number of generated surrogates.
6. Run the Adaptive Sampling design study.
  During the design study run, the surrogates are automatically created based on the Adaptive Sampling Settings you specified before. An example is shown below:
  
  You obtain one surrogate for each response. The adaptive sampling algorithm specified the surrogate type Kriging out of the two possible types.
7. To assess the prediction accuracy of the computed surrogates, investigate relevent data set and create plots as described in Assessing Surrogate Accuracy.
To generate the surrogates with a DOE study:
1. Create a design study.
2. Select the [desisgn study] node and set Study Type to DOE.
3. Select the DOE Settings node and set DOE Type to Latin Hypercube Sampling.
  The LHS algorithm promotes an even distribution of design points over the specified design space. This approach improves the accuracy of the computed surrogate approximation.
4. Select the Input Parameters node and add input parameters.
5. Select the Responses node and add design responses.
6. Right-click the [design study] > Surrogates node and select New Surrogate.
7. Select the Surrogates node and keep the Auto Compute activated.
8. Select the Surrogates > [surrogate] node and set its properties. For more details, refer to Surrogate Reference.
  A surrogate uses all the input parameters specified in the design study and provides approximation only for one response. This response you set in the Response of the [surrogate] node.
  After setting up surrogates, you can open up the Surrogate Table by right-clicking Surrogates node and selecting Edit Surrogate to multi-editing the surrogate settings. See also: Edit Surrogates.
  Alternatively, you can also create a surrogate for a certain response by right-clicking the [design study] > Responses > [response] node and selecting Create Surrogate.
9. Repeat the previous step for each response which you wish to analyse.
10. Run the DOE design study.
  The surrogates are automatically computed during the design study run.
11. To assess the prediction accuracy of the computed surrogates, investigate relevent data set and create plots as described in Assessing Surrogate Accuracy.

After evaluating the simulation-based design study in which the surrogates are created and computed, you either export surrogates to other tools or run a surrogated-based design study in Design Manager.

To export surrogates to other tools:
1. To export surrogates as Functional Mockup Units (FMUs), refer to Exporting Surrogates to Excel using VB Files.
2. To export surrogates as Visual Basic for Application Code Files (*.bas) in Excel, refer to Exporting Surrogates to Excel using VB Files.

In Design Manager, the most common applications of surrogated-based design studies are Robustness and Reliability studies due to the large number of designs.

To reflect the distribution of the responses, a Robustness and Reliability study often requires 2000 designs. Surrogate models are necessary to replace simulations

To set up a Robustness and Reliability study using surrogates:
1. Create a Robustness and Reliability study.
2. Select the [robustness and reliability study] node and set the Evaluation Method to Surrogates.
3. For input parameters, use values from the best design as Baseline Values and define their Standard Deviation (the default value is 10%). With those two values, normal distributions of the input parameters are defined. Statistical samples of the input parameters are used in numerous design studies.
4. Select the [robustness and reliability study] > Robustness and Reliability Settings node and set Designs to Run to a sufficiently large number, for example, 2000. The number should be sufficient to cover the normal distributions of the input parameters.
5. Select the [robustness and reliability study] > Robustness and Reliability Settings node and specify Sampling Method
  You are advised to apply the method Monte-Carlo if Designs to Run is set to a number over 10000.
6. Select the [robustness and reliability study] > Input Parameters node and add the input parameters used in the previous DOE study.
7. Select the [robustness and reliability study] > Responses node and add the responses whose surrogates are created in the previous DOE study.
8. Run the Robustness and Reliability design study.
9. Right-click the Design Studies > [robustness and reliability study] node and select Create Plot > Probability Distribution.
  
  In the probability distribution plot, you can compare the expected distribution of responses against a normal distribution. If the bars fall below the normal distribution of design responses, the baseline values of the input parameters are considered to be robust. Each response requires one individual probability distribution plot.