Generate new designs

With GeomAI, you generate new geometries based on existing ones. The principle is that:

1. You provide a dataset of geometries.

2. The AI model is trained to find a compressed representation of those geometries (the latent space).

3. Once the representation is computed, the model can generate new geometries by working in this compressed representation space.

Note

GeomAI is designed to generate geometries. It does not use physical response data (forces, flow fields, pressure, etc.) and cannot predict them. If your goal is to predict physical responses from geometries, use SimAI instead.

Key concepts

• A project holds your training data. The set of geometries the model learns from.

• A workspace holds a trained model and its predictions. One project can have multiple workspaces, each corresponding to a different build configuration.

The steps to follow are:

1. Prepare and validate your geometries.

2. Upload them as training data and assign them to a project.

3. Configure and train your AI model.

4. Evaluate the model quality.

5. Generate new designs and integrate them into your workflow.

For practical examples and scripts to execute yourself, see Examples. Follow the user guide to have more explanation on each step.

• Data preparation
  • Training data as design language
  • Geometry requirements
  • How to check and fix your geometries
  • Uploading your geometries
  • How much training data do you need?
• Model setup and training goals
  • Know your use case
  • Understanding reconstruction vs. interpolation
  • Build presets
  • Number of epochs
  • Number of latent parameters
  • Configuration example
• Evaluation and workflow integration
  • Evaluate your model with the Model Evaluation Report
  • Qualitative evaluation with sampling
  • Working with a reduced latent space
  • Troubleshooting and improving your model
  • Workflow integration