License checkout for prediction outside Altair GUI products, like HyperMesh or
Inspire, will now draw 150 Altair Units.
New Features
New Architecture Introduced - Transformer Neural Simulator (TNS)
PhysicsAI now has a new AI training method - the Transformer Neural
Simulator (TNS). This is in addition to the existing method, the Graph
Context Neural Simulator (GCNS). Some key differences between the two:
Usually, TNS should predict smoother contours than GCNS.
TNS is less sensitive to variation in mesh sizes.
On a GPU, TNS is faster than GCNS while on a CPU, it is
typically the opposite.
Natively Reading Simulation Entities (Thickness and Material IDs)
PhysicsAI can now extract thicknesses and material IDs if there are
solver decks available along with the training files. These decks should
be in the same location and have the same name as the corresponding
solver file. For example, if a file called
Ibeam.h3d has a file accompanying
Ibeam.fem, then the thicknesses and material
IDs can be parsed by selecting Extract Simulation
Properties.
Figure 1. The list of supported solver decks include: Optistruct, Radioss,
LS-DYNA, Nastran, Abaqus, and ANSYS.
Enhancements
Mesh Alignment for Translational Invariance
PhysicsAI is sensitive to meshes being translated and rotated in space.
Earlier, you had to manually orient the meshes to eliminate this source
of noise. Now, using the Mesh Alignment feature during model training,
the meshes can be adjusted during training, testing, and prediction. The
meshes are aligned such that the center of gravity is coincident at a
common point. This option can only correct translational variances and
not rotational variances.
Figure 2.
Figure 3.
Enhanced Dataset Visualization and Outlier Detection
Previously, you had to utilize other tools, such as HyperStudy, to
curate the data and identify outliers. Now, outliers in the dataset are
identified based on the you select. A Z-distribution is fitted to the
data and points that fall in the 3-sigma tails are highlighted as
outliers.
Figure 4.
Similarity Score in the PhysicsAI Connection in HyperStudy
You can now access the Similarity score for a prediction made using a
physicsAI model in HyperStudy. This can be a useful response to qualify
results based on expected accuracy. For example, using it as a
constraint to reject predictions with low Similarity scores. The
Similarity score is automatically added as a response if it is available
in the selected physicsAI model.
Figure 5.
Known Issues
Predictions made in the Radioss solver profile within HyperMesh may not
function correctly when model features are used (like thickness or
material). This issue can be avoided by using another solver profile, such
as OptiStruct.
Resolved Issues
Eroded elements are now hidden during visualization. Previously, eroded
elements resulted in an exploded mesh resulting in confusing visualization.
Now, elements are hidden after the point of erosion.
Figure 6.
Figure 7.
Previously, errors were encountered if the training data contained parts
with no results. Now, physicsAI automatically excludes such parts and the
training can continue.
