13 Frequently Asked Questions About Altair physicsAI
Altair EmployeeAltair physicsAI has brought advanced geometric deep learning to everyday CAE users. Get to know these basic facts before you get started.
HyperWorks 2022.3 contained the first release of our new physicsAI technology for fast physics predictions using historical simulation data. Since then, we have been hard at work listening to user feedback and improving. HyperWorks 2023 focuses on increasing usability and broadening the applications of physicsAI. In anticipation of the HyperWorks 2023 release, I want to share answers to some of the most asked questions about physicsAI. I’ve broken down the questions into three categories: (1) Data and Formats, (2) Computing and Resources, and (3) Accuracy and Quality.
Data and Formats
Q. What file formats are supported for training data?
A. The file reading technology is common through the HyperWorks ecosystem, for example HyperView.
Q. Are transient simulations supported?
A. Yes, both transient and static simulations are supported.
Q. How much data do I need?
A. Data is extracted from simulation results files. The number of result files required to get good quality predictions will vary between projects. Depending on the complexity of the physics and the amount of variation in the data, some applications may require only a handful of results while others require dozens or even hundreds. Additionally, sufficient quality is itself a subjective assessment. As a general guideline, it is recommended to train with at least 10 results before assessing tests of predictive quality.
Q. Do meshes need to have the same number of element/nodes?
A. No, it does not require equivalent meshes. The meshes do not even need to be topologically equivalent, as seen in the image below.
Q. How much can the design data vary?
A. There is no specific limit on the allowed variation in the training data. But it is informative to keep two considerations in mind. First, the training data should be representative of the new designs on which predictions will be made. Second, datasets with higher variability require a correspondingly larger number of training examples to maintain quality. When predicting, the confidence score can be used to quantify how similar the new design is to the training data.
Computing and Resources
Q. Is a GPU required?
A. No. A CPU can be used to train a model, but it will be slower than GPU.
Q. What GPUs are supported?
A. To train with a GPU you must install CUDA toolkit 11.8 and cuDNN 8.7. This requires an NVIDIA GPU of at least a Pascal microarchitecture, that is compute capability > 6.0.
Q. What effect does GPU and dataset size have on training times?
A. The quality of computing resources will affect training time. Better hardware can improve runtimes but improvements from CPU to GPU can be substantial. Regardless of hardware, training time is approximately linearly proportional to dataset size. See the chart below for a representative example:
|
| 10 result files | 50 result files | 100 result files |
| Laptop CPU | 44 m | 3 h 35 m | 7 h 54 m |
| HPC CPU | 34 m | 2 h 46 m | 5 h 10 m |
| GPU | 3 m | 16m | 33 m |
Q. Can I train on a an HPC?
A. Yes, more information is available in the documentation
Accuracy and Quality
Q. Is it accurate?
A. In general, the accuracy will improve with the amount of data, expressivity in the model (e.g. network width and depth), and allotted training time. However, practical considerations impose finite limits on these quantities. Quality assessment of a trained model is the final step in a standard training process. This is done by testing the predictions against known values, for example by using a test dataset's MAE metrics.
Q. What is a good MAE?
A. MAE is the mean absolute error. MAE can be interpreted as an error measure of a prediction. As an example, consider a model that predicts displacement with an MAE of 4 mm. You can interpret this as any given prediction may be inaccurate, on average, by 4mm. This may be significant if the predicted displacement of engineering interest is only 5 mm yet less consequential if a typical value is 500 mm.
Q. What training settings should I use?
A. Every project is different. The default settings are a good place to begin, however the best practice is to tune the settings to achieve a sufficiently high-quality model. The workflows permit the repeated training of models on the same dataset to compare the outcomes across different settings. These experiments may provide empirical evidence that similar projects may achieve the best results with similar settings.
Q. Can a trained model replace a solver?
A. Yes and no. Models are designed to act as a fast approximation of a solver, so in general, we do not expect solver-level accuracy. They are typically one to three orders of magnitude faster than a solver. This can be useful, even without achieving solver-level accuracy, because it allows you to rapidly explore new design concepts. The final design should always be verified with a traditional solver. That said, models can be trained to be quite accurate given the proper training data and settings.
Final Thoughts
The technology powering physicsAI will continue to expand. This isn't just a far off future vision; the tools are already embedded in today's software. It really is that simple! If you haven't already done so, get your hands on the latest version of HyperWorks and see how easy it really is for anyone to begin with AI-powered design.
Comments
-
Hi,
I am getting error "
Missing results found in input data - Verify that requested results are available for all nodes and elements"
while training the model.
My model does have stress output for all 2d elements but still the error of missing results.
May I know the reason?
Regards,
Mayur
0 -
Hello,
Thanks for your interest in physicsAI. The most likely cause is that there are additional elements beyond 2d that don't have the stress. For example, a rigid. Can you share which version you are running? Newer versions of physicsAI should be more forgiving. The current release, as of this response, is 2023.1.
We have worked around the issue in the older versions by removing the affected elements from the result file using a tool such as HVtrans.
If the problem persists, I recommend contacting Altair support and we can look into the specifics.
0