Altair PhysicsAI 2025.1 Release Notes

Announcements

PhysicsAI is now spelled with a capital P; physicsAI is no longer in use.

New Features

New Architecture: Shape Encoding Regressor
The Shape Encoding Regressor (SER) architecture uses the shape encodings of the training samples to fit multiple regressions. The predictions are restricted to scalar KPIs and vector curves (no contour results).
Figure 1.


SER requires you to select Extract Solid Faces during dataset creation.
SER does not support transfer learning.
In Table 1, the check marks indicate supported capabilities.
Table 1. Comparison of the Three Architectures
GCNS TNS SER
Contour Output ✓ ✓
KPI or Curve Output ✓ ✓ ✓
Mesh Input ✓ ✓ ✓
CAD Input ✓ ✓
Custom Inputs ✓ ✓
Contour Smoothness ✓ NA
Training Time - +
Training Stability - +
GPU Memory + - NA
The best use cases for the three architectures are as follows:
  • GCNS is best for general use.
  • TNS is best for predictions on CAD.
  • SER is best for KPI predictions without a GPU.
User Input Validation Fraction
Validation datasets are useful in preventing overfitting on large datasets. However, on small datasets it can lead to the selection of a suboptimal epoch as the best model configuration.
The default value is 0.15 if there are more than 50 samples and 0.0 for 50 samples or less.
Figure 2.


Enhancements

Element Corner Data Support
Element corner data can now be read into PhysicsAI.
Similarity Score for Curve Prediction
Similarity score is now available for curve predictions.
Figure 3.


Enhanced Data Curation Capabilities
The data curation features (part of dataset creation) have been significantly enhanced. The updates can be divided into two categories: data inspection and data management.
Data Inspection
Updates to the model summary and results metrics table: outliers are now automatically identified and highlighted in orange. The outlying columns are also displayed. The outlying KPI value is highlighted in bold font. You can select or remove additional responses by interacting with the tree shown on the right side of Figure 4.
Figure 4.


Plotting metrics: this is an additional visualization tool which can be accessed by switching to the Plot tab. This feature is similar to parallel plots in HyperStudy. The distribution of the sample KPIs and outliers can be graphically identified, which complements the tabular view of the model and results summary table.
Figure 5.


Figure 6.


Rotational outlier identification: rotational outliers can now be identified by using the dimensions and center location of the bounding boxes of the samples or datapoints.
Figure 7.


Data Management
Datasets can now be edited to add more samples or remove outliers.
Figure 8.


Figure 9.


Samples can be moved between datasets.
Figure 10.


Entire datasets can be cloned to preserve originals without re-extraction.
Figure 11.


KPI and Curve Prediction Visualization in HyperMesh CFD
The PhysicsAI ribbon inside HyperMesh CFD can now display the predicted curves and KPI.
Figure 12.


Faster Predictions
Prediction time on new designs now takes 50% less time compared to earlier versions.
Figure 13.


Resolved Issues

  • Memory issues have been fixed that used to cause a varying memory requirement for dataset creation. Now, a constant memory will be consumed during dataset creation.

Known Issues

  • On Linux, PhysicsAI 2025.1 requires a mandatory update of cuDNN to 8.9.7. Not updating cuDNN may lead to inadvertent training with a CPU despite expecting training with a GPU.
  • Projects that contain a model which has an error state will not open correctly on the first try. Such projects will open correctly if reopened a second time as part of the same session.