3D-to-Scalar

Modified on Fri, 16 Dec, 2022 at 2:29 PM

Description

3D-to-Scalar model is a new 3D Deep Learning model which links unstructured 3D CAD designs data directly to scalar quantities of interest.

It will allow you to drop a new 3D CAD design into Monolith and get an instant prediction of its performance or quality under new operating conditions.

Application

3D-to-Scalar is suitable for users with unstructured 3D CAD design and scalar quantities of interest who wish to accurately predict the performance of a specific input mesh.

This means the performance data is included in the loss function of the model as it learns to encode 3D data. For many use cases, it will therefore provide users with more accurate predictions. These models will only be used for validation use cases (predicting the performance of an existing design) - not to generate new, performance-optimised designs (which is the role of the Autoencoder).

How to use

The user workflow is straight-forward and involves training one model.

Step 1

Import any 3D models that are supported on the platform. Learn more about the types of 3D data supported.

Step 2

Import tabular data: scalar performance metrics and operating conditions, this should include both test inputs and outputs.

Step 3

The tabular dataset needs to be attached using the Attach Parameters to 3D, note that an ID column must be available to map the 3D object names to the scalar performance metrics. In this step you will need to select the parameters to attach, these will then be available as inputs and outputs in Step 4 & 5.

Step 4

Train a 3D-to-Scalar Model, this step is available in Models for 3D Data. You will need to select inputs and outputs for model training from a list that is populated based on the attached parameters from Step 3. Advanced options in this step give the user access to three hyperparameters: learning rate, early stopping and N Samples.

Learning rate	This changes the step size at each iteration, increasing this may result in faster training but increases the risk of overshooting.
Early stopping	This stops the model training if no model improvement is seen in a certain number of epochs. The user can change the number of epochs from 10, 15, or 20.
N samples	The number of sample points from each model, increasing this improves the model by increasing its complexity but can significantly increase training time.

Step 5A

To get a visual understanding of each individual prediction, select View 3D-to-Scalar Predictions within the Apply manipulator, to visualise individual predictions. This will allow you to change the model inputs and view the impact these have on the chosen outputs.

Step 5B

3D Dataset Prediction, to tabulate predictions.

What are the main differences between autoencoders and 3d-to-Scalar?

3D-to-Scalar

Model learns to accurately predict scalar performance metrics
Only for Prediction: Assessing the performance of existing designs
Fast training time and simplified user workflow
No pre-processing steps diluting raw data

Autoencoder

Model learns to accurately parameterise geometry (a second, downstream model is needed for performance predictions)
For Prediction and Optimisation: Generating performance-optimised designs
Slower training time for unstructured data
Voxelisation dilutes raw 3D data into a voxel grid