Guidelines for PhysicsAI Usage

Data Selection

PhysicsAI correlates geometric features and results, so input data should contain sufficient data points and be varied enough to be a good representative of the design space.

• Ensure that the training dataset is sufficiently large and varied.
• Small training sets lead to very small validation sets, risking overfitting.
• Curate data by removing outliers.
• Ensure that data is rotationally and translationally aligned.
• Ensure that data does not include overlapping faces featuring multiple results like baffles or inlets/outlets of porous media.

Data Consistency

Consistent result data

Ensure that results are available on all nodes in the model. This ensures that learning algorithms do not assume any data and output inaccurate results.

Consistent geometry configuration

Training data should have similar geometries so that the training algorithm learns the impact of small feature changes. For example, it is not recommended to train sedan and pickup truck models together.

Consistent physics and boundary conditions

All results should correspond to the same physics and boundary conditions as these are not considered during training by default.

• Segregate simulations with similar physics and boundary conditions.
• Custom scripts can be created to transfer physics data and boundary conditions for training.

Data Decimation

• If training with the SET method to predict KPIs, decimation is not required.
• If training with the TNS method to predict surface fields, decimation is required.
• Current implementation supports single GPU training, restricting the model parameters and model size that can be trained with. Current capability supports surface field prediction but does not support volumetric field data training for CFD models. For surface field prediction, you should decimate the model for optimized training.
• The Decimate tool enables you to decimate results files (.case, .vtk, .h3d formats) and to save as .h3d files.
• Ensure that results are available on all nodes, even they are zero.
• The Decimate tool outputs .h3d files with coarse triangulated surface fields.
• Large triangles may be created in flat areas, so it is recommended to choose an appropriate reduction factor and feature angle.
  • If reduction is not possible, the alternative is to apply a smaller percentage of sampling points/epoch when using the TNS method and to reduce the hyperparameters.
• If the input surface mesh has connectivity issues, there will be holes in the decimated results. Ensure that inputs do not have any artifacts.
• The Decimate tool allows selected parts to be skipped during processing. For external flow, ensure that inlet and outlet interior faces are skipped to avoid issues in training.

Transfer Learning pulls weights from an existing model and updates the current model accordingly.

Augment

Keeps the existing weights as is and appends additional training data. Introduces a smaller chunk of new training data compared to the baseline model training data.

Initialize

Introduces any amount of training data.