Browsers provide a structured view of model data, which you can use to review, modify, create, and manage
the contents of a model. In addition to visualization, browsers offer features like search, filtering, and sorting,
which enhance your ability to navigate and interact with the model data.
FE geometry is topology on top of mesh, meaning CAD and mesh exist as a single entity. The purpose of FE geometry
is to add vertices, edges, surfaces, and solids on FE models which have no CAD geometry.
Tools and workflows that are dedicated to rapidly creating new parts for specific use cases, or amending existing
parts. The current capabilities are focused on stiffening parts.
Use PhysicsAI to build fast predictive models from CAE data. PhysicsAI can be trained on data with any physics or
remeshing and without design variables.
Explore, organize and manage your personal data, collaborate in teams, and connect to other data sources, such as
corporate PLM systems to access CAD data or publish simulation data.
Use PhysicsAI to build fast predictive models from CAE data. PhysicsAI can be trained on data with any physics or
remeshing and without design variables.
From the PhysicsAI ribbon, select the Train
an ML Model tool.
The Train Model dialog opens.
Define the training details and click Train.
Enter a model name.
For Training Data, select the created or existing Train Dataset.
Select Inputs and Outputs.
Specify the Epochs, the number of times the data needs to train.
The Model Training dialog
opens.
Review the status in the Status column.
Tip: Once the status changes to Running, you can view the training
logs by clicking Show Log.
Optional: Click Loss Curve to view the training of a validation
loss curve.
The curves are useful to visualize the progress of the training process. In a
well fit model, the training and validation losses become nearly identical. If
validation never approaches the training loss, this is indicative of
underfitting; increased training time can leave files d to improved model
performance. A validation loss that approaches the training loss but diverges
higher likely indicates overfitting; the point of low validation loss is the
ideal model to avoid loss of generalization.
Note: The validation curve only
appears if there are at least 15 samples.
Train Remotely on an HPC
Train a PhysicsAI model on a remotely on a different machine than the one running the
PhysicsAI GUI.
To train models remotely, you
will need:
Access to a remote machine with the Engineering Data Science (EDS) application
installed
PuTTY installed on your local machine
A training script (details below)
A mapped drive which can be accessed by both your local machine and the HPC.
This is required so that your locally created datasets are visible to the HPC
during training.
A common reason for remote training is to harness an HPC with a GPU, which can
accelerate training significantly.
Install the EDS application from the AltairOne Marketplace to your HPC.
Create an SSH connection.
Launch PuTTY on your local machine and connect to the HPC via
SSH.
Save the connection, for example:
my_physicsAI_hpc.
Important: If you are required to enter a password while logging in
via PuTTY, you will need to setup RSA keys before continuing. Once you can
login via PuTTY without entering a password, you may continue.
Write a training script for your HPC using the following template.
This example uses the qsub command from PBS on Windows. For Windows, the script should have the
.bat extension. If your local machine is running Linux,
you will need to write a script with equivalent functionality on Linux with a
.sh extension.
@echo off
SETLOCAL
REM ------------------------------------------------------------------------------
REM Copyright (c) 2021 - 2021 Altair Engineering Inc. All Rights Reserved
REM Contains trade secrets of Altair Engineering, Inc. Copyright notice
REM does not imply publication. Decompilation or disassembly of this
REM software is strictly prohibited.
REM ------------------------------------------------------------------------------
REM ------------------------------------------------------------------------------
REM USER SETTINGS
REM ------------------------------------------------------------------------------
REM HPC Setup
set sess=<<HOST NAME>>
set user=<<USER NAME>>
REM PBS Requests
set pbs_requests=-q a100 -N physicsAI_shape -j oe -l select=1:ncpus=8:mem=257940mb:ngpus=1
REM Windows -> Unix Mapping
set win_map=\\<<SERVER IP>>\data\ds
set unix_map=/data/ds
REM PhysicsAI Installation Settings
set install_loc=<<HW INSTALL LOCATION>>hwdesktop/hw/eds/bin/linux64/edspy.sh
REM PBS Install Location
set sub_cmd=/altair/pbsworks/pbs/exec/bin/qsub
REM ------------------------------------------------------------------------------
REM SCRIPT START
REM ------------------------------------------------------------------------------
REM Get whole input line
set line=%*
REM Map windows paths to unix
set unixmap=%win_map%=%unix_map%
set unixmap=%unixmap:\=/%
set submit_line=%line:\=/%
setlocal EnableExtensions EnableDelayedExpansion
set submit_line=!submit_line:%unixmap%!
echo %line%
echo ---UNIXMAP---
IF ["%unixmap%"] == [""] GOTO :RUN
setlocal EnableExtensions EnableDelayedExpansion
:RUN
echo ---PLINK SETTINGS---
echo user = %user%
echo sess = %sess%
echo ---PAI-SHAPE SETTINGS--
echo install_loc = %install_loc%
echo submit_line = %submit_line%
REM -----------------
REM Run QSUB
REM -----------------
set qsub_command_string='%install_loc% %submit_line%'
REM Write File for submission
echo plink -load %sess% -l %user% -batch "echo %qsub_command_string% > pai_qsub.txt"
plink -load %sess% -l %user% -batch "echo %qsub_command_string% > pai_qsub.txt"
REM Submit file
plink -load %sess% -l %user% -batch "%sub_cmd% %pbs_requests% pai_qsub.txt"
echo plink -load %sess% -l %user% -batch "qsub %pbs_requests% pai_qsub.txt"
ENDLOCAL
Register the training script.
From the PhysicsAI ribbon, select the
Train an ML Model tool.
The Train Model dialog opens.
For Training Script, select Register Training
Script.
The Add Training Script dialog
opens.
Click and browse and
select your training script.
Enter a name for your script and click OK.
Close the Train Model dialog.
Your preferences are saved and the
physicsai_solver_prefs.json file in your user directory
has been updated.
Once these tools are installed, the GPU will be used by default for both
training and predicting. You can verify this in the Task Manager by enabling the
cuda graph in the GPU Performance tab.
To use the CPU again, set CUDA_VISIBLE_DEVICES=-1 as an
environment variable.