A user shares their experience with the computational expense of generating training datasets from physics simulations, such as fluid flow and structural fields. They highlight that the primary challenge is not the model architecture but the difficulty of building a large and diverse dataset when each sample is costly to produce.
The post seeks community input on several key aspects of managing simulation data:
- Sampling strategies for expensive cases, including Latin hypercube, active learning, and adaptive sampling.
- Data representation formats for field data on meshes or grids, such as point clouds, voxel grids, graphs, or resampled uniform arrays.
- Augmentation techniques like symmetry, rotation, and superposition, and their potential to break underlying physics.
- The use of the model itself in active learning to decide which new expensive samples to generate next.
The author is specifically interested in real-world experiences dealing with the "few, expensive samples" regime rather than standard big-data assumptions.