Use SER Architecture for Training on KPIs

Tutorial Level: Advanced In this tutorial, you will use the Shape Encoding Regressor (SER) architecture to predict KPIs.

The SER architecture operates on shape encodings of input samples and can only predict KPI/vectors (no contour or field results).

In this tutorial, you will:
  • Create and curate a dataset to remove outliers.
  • Train a PhysicsAI model to predict a KPI using the SER architecture.
Before you begin, copy the file(s) used in this tutorial to your working directory.
Note: Unzip the project IBeam_dataset.zip and inspect the contents.
  • run_000XX_Ibeam.h3d files
  • run_000XX_Ibeam.json files containing the KPIs

Create and Curate a Data Set

  1. Open HyperMesh.
  2. From the menu bar, click View > Ribbons > PhysicsAI to open the PhysicsAI ribbon.
  3. Create a project.
    1. From the PhysicsAI ribbon, select the Create Project tool.
      Figure 1.


      The Create Project dialog opens.
    2. For Project Name, enter SER_Tutorial.
    3. For Location, click Choose and select a save location for the project.
      Note: The save location for the project contains all files created by PhysicsAI, but the original files used for training do not need to reside in the project folder.
    4. Click OK.
  4. Create a database called Ibeam_full using all the .h3d files.
    1. From the PhysicsAI ribbon, select the Create Dataset tool.
      Figure 2.


      The Create Dataset dialog opens.
    2. For Dataset Name, enter Ibeam_full.
    3. For File System, click and navigate to the IBeam_dataset folder.
    4. Select and transfer all of the .h3d files.
    5. Select the Enable train test split check box and verify the Train % is set to 80.
      Note: Enabling the 80:20 split will generate two datasets: Ibeam_full_train with 26 files and Ibeam_full_test with seven files.
    6. Verify the Extract solid faces check box is selected.
      Important: This is required for the SER architecture.
    7. Click OK.
      Figure 3.


    The dataset is extracted and the Datasets dialog opens.
    Figure 4.


  5. Remove outliers in the training dataset.
    1. Select Ibeam_full_train.
    2. Select the first orange highlighted row, row two, and click to remove the row from the dataset.
    3. Repeat step 5.b to remove all other rows that are outliers.
    The curated dataset no longer has any outliers.
    Figure 5.


Train Model Using SER Architecture

In this step, you will train a PhysicsAI model using the SER architecture.

  1. Create a new model called Max_disp_SER.
    1. From the PhysicsAI ribbon, select the Train an ML Model tool.
      Figure 6.


      The Train Model dialog opens.
    2. For Model Name, enter Max_disp_SER.
    3. For Training Data, verify Ibeam_full_train is selected.
    4. For Vector, select Max_disp_LC1 which is the response of interest corresponding to the maximum displacement in loadcase 1.
      Note: Since all the .h3d files used in the Ibeam_full_train database have a corresponding .json file, the KPIs from these files are displayed in the Vector Outputs drop-down menu.
    5. For method, select SER.
    6. Verify K-folds is set to 10.
      Note: This controls the number of random splits for the training and validation sets. This number should be sufficiently large to allow for a well performing model, but not too large to avoid overfitting. The larger the number of K-fold build, the longer the training time.
    7. Verify K-fold test fraction is set to 0.1.
      Note: This is equivalent to the validation fraction and control the percentage of data to be used for validation per K-fold build.
    8. Verify the PCA input check box is selected to calculate the principal component of the encodings.
      Note: Calculating the principal component of the encodings helps to fasten convergence. The PCA output can be enabled in case of vector (longer than 1) outputs.
    9. Click Train.
    Figure 7.


  2. Test the model.
    Note: SER is much faster than GCNS and TNS. Only the training loss is available for SER and there is no loss curve.
    1. From the PhysicsAI ribbon, select the Test ML Model tool.
      Figure 8.


      The Test Model dialog opens.
    2. For Models, select Max_disp_SER.
    3. For Datasets, select Ibeam_full_test.
      Figure 9.


    4. Click OK.
      The Model Testing dialog opens.
  3. Select a single data point and click Display File.
    Figure 10.