Frequently Asked Questions
This section provides quick responses to typical and frequently asked questions regarding physicsAI.
Data and Formats
- What file formats can be used for training physicsAI models?
- The file reading technology is common throughout the ecosystem, for example HyperView. For more information, see Coverage.
- Are transient simulations supported?
- Yes, both transient and static simulation are supported.
- How much data is needed?
- The number of result files required to get good quality predictions will vary between projects, and sufficient quality is itself subjective. Some application problems may require only a handful of results, while others require dozens or even hundreds. As a general guideline, it is recommended to train with at least 10 results before assessing tests of predictive quality.
- Do meshes need to have the same number of element/nodes?
- No, the meshes do not require equivalent mesh. The meshes do not even need to be topologically equivalent.
- How much can the design data vary?
- 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 type of data 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 design is to the training data.
- Do I need to include simulation entities like thickness, materials, loads, boundary conditions, and so on in my physicsAI training?
- Not necessarily. These entities only need to be included if they vary between training samples. For example, if thicknesses are constant across all elements in all meshes, then there will be no benefit to including them in training. The same is true for other entities; only data with variation is important. However, if thickness, or other entities, vary from one training sample to another, then it is necessary that you include them in physicsAI training to obtain quality predictions.
- How do I include simulation entities like thickness, materials, loads, boundary conditions, and so on in my phyiscsAI training?
- PhysicsAI can support simulation entities with custom features. The preferred option is native support; see this list of natively supported entities. The second best option is to use a custom hook (See Custom Inputs: Global Hook and Custom Inputs: Nodal Hook for more information) that provides flexibility to support virtually any entity via reusable and sharable code. The third choice is the predefined nodal or global hook that provides a simple entry point for working with data via a simple human readable format.
- What entities are natively supported?
- The following entities are automatically parsed by physicsAI.
Version Solvers Entities 2025 OptiStruct Radioss
LS-DYNA
Nastran
Abaqus
ANSYS
Thickness Restriction: Works with thickness assigned to properties. Thickness assigned directly to elements is not supported.Material ID
- How does physicsAI handle your data?
- You control your own data. You can install physicsAI in your own system like conventional CAE software. You can create projects on your file systems, and any data created by physicsAI remains in this location and is not moved outside of the project's context. PhysicsAI supports offloading the calculations to HPC or cloud environments, but only by customization.
- What are the different AI architectures/methods available in physicsAI?
- In physicsAI, you can choose from two options: the Graph Context Neural Simulator (GCNS), which is the default option, and the Transformer Neural Simulator (TNS).
- What are the differences between the Graph Context Neural Simulator (GCNS) and the Transformer Neural Simulator (TNS)?
- The GCNS is based on graph neural networks. This was the only method
available in HyperMesh 2024.1 and earlier
versions. The TNS is based on the transformer architecture and is available
in HyperMesh version 2025.0 and later. 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. On a CPU, it is typically the opposite.
- Can physicsAI handle training files with missing results on some parts?
- Yes, in physicsAI 2025.0 or later, parts without any results are automatically detected. Using the label IDs, these parts are excluded during training as well as prediction.
- Can physicsAI handle element erosion?
- Yes, eroded nodes (nodes with no element connections) are ignored.
- Do all the meshes in the training data need to be aligned? Does physicsAI have any utilities for this?
- PhysicsAI is sensitive to location and orientation of training data. In physicsAI 2025.0, meshes can be translated using the center of gravity to the origin of the reference coordinate system using the Mesh Alignment Feature. For details, see Train Models.
Computing and Resources
- Is a GPU required?
- A CPU can be used to train a model, but it will be slower than a GPU.
- What GPUs are supported?
- To train with a GPU, you must install CUDA toolkit 11.8 and cuDNN 8.7. This requires a NVIDIA GPU that has a compute capability > 6.0. See the NVIDIA website for more information on the compute capability of specific hardware.
- How do I enable NVIDIA GPUs for physicsAI training?
-
- Verify the GPU is supported.
This requires a NVIDIA GPU that has a compute capability > 6.0. See the NVIDIA website for more information on the compute capability of specific hardware.
- Install CUDA toolkit 11.8 and cuDNN 8.7.
- Update the "path" environment variable to include the following
directories (this may be different on your machine). For example, on a Windows PC, the following paths need to be added.
- C:\Program Files\NVIDIA\CUDNN\v8.7\bin
- C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin
- C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\libnvvp
- Relaunch HyperMesh.
- Verify the GPU is supported.
- What effect does GPU and dataset size have on training times?
- The quality of computing resources will affect training time. Better
hardware can improve run times, but improvements from CPU to GPU can be
substantial. Regardless, training time is linearly proportional to dataset
size. See the table below for a representative example:
Hardware 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 - Can I train on an HPC?
- Yes, more information is available here.
Accuracy and Quality
- Is physicsAI accurate?
- In general, the accuracy of a physicsAI model will improve with the amount of data, expressivity in the model (for example, width and depth), and allotted training time. However, practical considerations impose finite limits on these quantities. Assessing the quality of a trained model, for example by testing the model’s MAE metric against known values, is a standard step in the process of training models.
- What is a good MAE?
- MAE is the mean absolute error. MAE can be interpreted as an error measure of a prediction. For example, consider a model that predicts displacement with an MAE of 4 mm. This means 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.
- What training settings should be used?
- Every project is different. The default settings are good places to begin, but the best practice is to tune the settings to achieve a sufficiently high-quality model. The physicsAI 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.
- Can a physicsAI trained model replace a solver?
- Yes and no. PhysicsAI models are designed to act as fast approximations of a solver, so in general, we do not expect solver-level accuracy. It is typically one to two orders of magnitude faster than a solver. This can be useful, even without achieving solver-level accuracy, because it allows you to rapidly analyze new design concepts. The final design should always be verified with a traditional solver. That said, physicsAI models can be trained to be quite accurate given the proper training data and settings.