Machine Learning with the Geometry Power Tool

This page describes the machine learning tools that are part of Cubit's Geometry Power Tool for tet mesh quality prediction and part classification.

Background:

The Geometry Power Tool in Cubit provides a series of diagnostic checks on your model used to defeature or simplify a CAD model prior to meshing. Clicking the Analyze button will perform the selected diagnostic tests and display an expanding tree listing geometric entities that are identified by each test. Once identified, suggested solutions can be easily previewed and executed.

This new capability enhances the Geometry Power Tool using state-of-the-art machine learning (ML) methods to predict meshing outcomes, suggest solutions as well as classify certain common part types.

Enable Machine Learning:

Use the following procedure to access the new ML capabilities:

Navigate to the Geometry Power Tool by clicking on the Geometry icon in the power tools window as shown in Figure 1.

Figure 1. Navigation to Power Tool and Options panel.
Select the Options... button shown in Figure 1.
The panel shown in Figure 2. should appear. Click the button labelled Load ML Models
Once ML models have loaded, additional diagnostics will be available for selection as shown in Figure 3. These include Tetmesh Poor Quality Predictions and Part Classification.

Figure 2. Options panel without ML models loaded

Figure 3. Options panel after loading ML models.

Geometry Analysis Tools

The following describe the additional diagnostics available when ML is enabled in the Geometry Power Tools.

Beams and Shells
Slot Surfaces
Tetmesh Poor Quality Predictions
Part Classification
Surface Classification

Beams and Shells

Geometry simplification is often required when modeling assemblies comprised primarily of thin volumes. Shell finite elements are often used rather than full 3D hex or tet elements. The task for analysts in this case is to reduce the 3D set of thin volumes to a set of connected sheet bodies where a mesh of triangles or quadrilaterals can be applied. This procedure can be managed with the Geometry Power Tool if the Beam and Shell diagnostic is selected. See Beams and Shells with the Geometry Power Tool for details.

Slot Surfaces

The Slot Surfaces diagnostic is a new addition to the geometry power tool for preparing slot surfaces in Electromagnetic (EM) modeling. A slot surface is used to identify potential pathways where EM radiation can potentially exit. This diagnostic will help manage slot surfaces and prepare them for the application of boundary conditions and analysis. This diagnostic utilizes Machine Learning (ML) to predict the most likely slot surfaces. See Slot Surface Preparation with the Geometry Power Tool for details.

Tetmesh Poor Quality Predictions

The solid model used as the basis for a tetrahedral mesh may contain small features or angles that can lead to poor mesh quality or very small elements. This diagnostic will provide a list of entities (vertices, curves, surfaces) that are predicted to result in poor quality tet elements sorted by their predicted mesh quality metric. This allows the user to quickly focus on regions of the model that will result in the worst mesh quality and apply geometry operations to improve the meshing outcome without having to mesh.

Three different mesh quality metrics are used for the basis of the predictions: Scaled Jacobian, In-radius and Deviation. Entities identified by these diagnostics will be sorted according to their ML-predicted metric starting with the worst quality entity. The three edit fields at the top of the Options panel control the quality limit. For example, a Scaled Jacobian Limit of 0.2 will identify all geometry entities predicted to have nearby tet elements whose Scaled Jacobian is less than 0.2. For our purposes, "nearby tets" are defined as those within two edge lengths of the geometric entity.

Poor Scaled Jacobian - Select this option if you would like to view mesh quality predictions based on Scaled Jacobian. Scaled Jacobian is a value between -1.0 and 1.0, where 1.0 is ideal and -1.0 is completely inverted. This value is a function of the angles at the vertices of the tet so that values less than or close to zero should be avoided for analysis.
Poor In-Radius - Select this option if you would like to view mesh quality predictions based on In-Radius. The in-radius is defined as the radius of the largest inscribed sphere in a tetrahedron. We are interested in a measure of how closely the tet elements meet the user-prescribed mesh size. Since in-radius is an absolute distance, we use the in-radius ratio defined as in-radius/ideal-inradius. The ideal in-radius is defined as the radius of an ideal tet (Scaled Jacobian = 1.0) whose edge lengths are precisely the user-specified mesh size. Therefore, a value of 1.0 meets exactly the mesh quality requirement, while a value less than 1.0 is smaller than prescribed. To determine the ideal in-radius current mesh size defined on the local volume will be used. if not set, then a target of 4*small curve threshold will be used.
Poor Deviation - Select this option if you would like to view mesh quality predictions based on Deviation. Deviation is defined as the distance from the midpoint of a triangular face at the boundary to its closest point on the geometry. A large value for deviation indicates that the mesh does not conform well or deviates significantly from the geometry. As deviation is also an absolute distance, we define the deviation ratio as deviation/mesh size. A value close to 0.0 for deviation ratio is better as it indicates the triangular faces very nearly match the geometry. A larger value closer to 1.0 indicates the local deviation from the geometry is on the order of the mesh size. To determine the deviation ratio, the current mesh size defined on the local volume will be used. If not set, then a target mesh size of 4*small curve threshold will be used.

Context Menus: To display all entities in a specific category, select the diagnostic category title and use the right click context menu to choose Draw. Expanding the list will display individual entities that can be visualized using standard draw tools such as Zoom To, Fly In, Draw Owning Volume, Draw with Neighbors, etc. When one or more entities are selected in the list, a context menu also provides access to selected command panel options. When multiple entities are selected, the context menu can be used to invoke command panels to operate on multiple entities simultaneously.

Part Classification

This diagnostic provides the ability to classify volumes according to several common part types. When selected, characteristic geometric features of each volume will be computed and ML methods will be used to determine the part classification. Each volume will be placed into a list based on its most probable categorization. Volumes in each classification will be ordered based upon a Confidence metric. A confidence of 1.0 indicates a 100% confidence of categorization. Confidence values closer to 0.5 are less likely to be categrized correctly, but ML indicated a higher probablility than other part types. The current part categories are illustrated below showing a few examples from each category. In addition to these Cubit's classify command can manage custom creation and editing of part categories.

Bolts
Nuts
Washers
Springs
Balls
Bearing Races
Pins
Gears
Other Parts

Note that the Other Parts category simply includes parts that cannot be categorized with sufficient confidence in any of the other categories.

Context Menus: To display all volumes in a specific category, select the part category title and use the right click context menu and choose Draw. Expanding the list will display individual volumes that can be visualized using standard draw tools such as Zoom To, Fly In, Draw Overlapping Volumes, Draw Nearby Volumes, etc. The context menu also provides access to command panel options such as Delete or Reduce. When multiple volumes are selected, the context menu can be used to invoke command panels to operate on multiple volumes simultaneously.

To view the volume features or the confidence values computed by ML, use the right click context menu when a part/volume is selected and choose either List ML Features or List ML Predictions menu option.

Surface Classification

This diagnostic provides the ability to classify surfaces. Surface categories provided with Cubit are currently limited to two categories: slots and non_slots. The limited categories are intended to be exemplars for additional user customized categires that can be defined using Cubit classify command. Similar to part classification, when selected, characteristic geometric features of each surface will be computed and ML methods will be used to determine its categorization. Each surface will be placed into a list based on its most probable categorization. Surfaces in each classification will be ordered based upon a Confidence metric. A confidence of 1.0 indicates a 100% confidence of categorization. Confidence values closer to 0.5 are less likely to be categrized correctly, but ML indicated a higher probablility than other part types.

Solution Predictions

Tetmesh Poor Quality Predictions

Once entities have been identified, custom solutions may be displayed by checking the Show Solutions check box above the list of entities. When checked, the Solution window will appear as shown in Figure 4. When an entity is selected in the upper window, an ordered list of solutions will appear in the Solution window. Each solution has an associated metric prediction so that the solution predicted by ML to provide the best possible meshing result will appear at the top of the list.

Figure 4. Geometry Power Tool showing Scaled Jacobian quality predictions as well as an ordered list of possible solutions.

For example, in Figure 4. Surface 1786 is selected which has a predicted minimum Scaled Jacobian metric at the surface of 0.0164. In most cases this metric would be unacceptable for analysis. To correct for this, the first solution in the list is a Replace Surface command that is predicted to result in a Scaled Jacobian of 0.2623. Selecting the solution will display a preview of the operation and double clicking will execute the solution. Other solutions may also be previewed and considered, however those with lower predictions would most likely be avoided.

Context Menus: A context menu is available when selecting a solution. This also provides an option for execution of the solution as well as additional visualization options. It also provides direct access to the appropriate command panel if further customization of the command is desired.

Controlling Solution Predictions: When selecting an entity from one of the metric categories (Scaled Jacobian, In-Radius, Deviation), the predicted value and ordering of the solutions will be based on the corresponding category. Note that solutions may also be generated for other diagnostics (ie. small curves, close loops, blend surfaces, etc.). When ML has been enabled, solutions will also be ordered these categories based upon the predicted tetmesh outcome. The metric used for ordering solutions can be customized using the Options panel shown in Figure 3. Use the Show Solution Predictions check box to toggle the use of ML-predictions to order solutions. When solution predictions are enabled, a drop-down selector will allow selection of one of the three metrics to prioritize solutions.

Part Classification

When selecting a volume from a classification category, selected solutions will also be displayed. Custom solutions based upon a specific part type are still in development and will expand in future versions of the geometry power tool. In particular, the reduce family of commands are available when one or more volumes classified as a "bolt" or "spring" are selected.