Generative design#
Learn about the parameters to set to generate a new design based on a trained model.
Latent parameters#
The latent parameters (latent_params) correspond to a list of numbers (floats) that represent the position of the geometry in the latent space.
You define this parameter to generate a geometry with a trained model.
The number of floats can be less than or equal to the nb_latent_param your model was requested with.
For more information, see Number of latent parameters.
Working with reduced latent spaces#
GeomAI expresses its latent space in a basis where dimensions are ordered by importance: the first dimensions capture the largest sources of variation in the training data, while later dimensions capture increasingly finer details. This means that for many workflows (exploration, interpolation, optimization), you can work with a small subset of the most informative dimensions rather than the full latent vector.
Choosing the right number of dimensions#
The Model Evaluation Report (MER) has a Parameter Importance section that shows how much each
latent dimension contributes to the variability of the training data. Use this to decide how many
dimensions to keep: if the first 10 dimensions account for most of the variance, working with
n=10 gives you a compact space with minimal loss of geometric fidelity.
Retrieving reduced latent codes#
The workspace.get_latent_parameters() method accepts an optional n parameter that truncates
each latent vector to its first n elements. The value of n must not exceed the number of
latent parameters the model was trained with.
# Full latent codes
full = workspace.get_latent_parameters()
# Only the 10 most important dimensions
reduced = workspace.get_latent_parameters(n=10)
When n is omitted, the full vectors are returned as before.
Generating geometries from reduced codes#
Reduced latent vectors can be passed directly to GeomAIPredictionConfiguration. The remaining
dimensions are automatically discarded, so no additional handling is required.
from ansys.simai.core.data.geomai.predictions import GeomAIPredictionConfiguration
vec = list(reduced.values())[0] # 10-element vector
config = GeomAIPredictionConfiguration(latent_params=vec)
prediction = client.geomai.predictions.run(config, workspace)
This is particularly useful for integration with optimization tools, where working in a lower-dimensional space reduces the search space and speeds up convergence.
Note
Reducing the number of latent dimensions discards the information carried by the removed dimensions. Geometries reconstructed from reduced codes are close approximations of the originals, not exact replicas. The fewer dimensions retained, the coarser the approximation.
Resolution#
The resolution parameter corresponds to a list of three integers defining the number of voxels along the X, Y, and Z axes.
The total number of voxels must not exceed 900^3 (that is, X × Y × Z ≤ 900^3). Exceeding this limit will result in an error.
Prediction time increases with the number of voxels. For the maximum resolution of 900^3, the prediction takes approximately 10 minutes (approximately 1 microsecond per voxel).
By default, this parameter is set automatically based on the training data used to build the model.
How does the default auto-resolution system work?
During training, each dataset is assigned a resolution derived from its mesh characteristics, particularly average edge lengths, so that reconstructed outputs preserve similar geometric fidelity.
At inference, the system finds the nearest training example in the latent space, and reuses its associated resolution.
As a result, the resolution is not fixed per model but varies per inference, it provides you with a data-driven default to ease user experience. You can still adjust it manually by using higher resolution for complex or precise geometries, and lower resolution for simple shapes or quick previews.