Generating and curating training datasets from simulation — how are people handling the data bottleneck for scientific ML?

Researchers working with scientific machine learning face a data bottleneck because generating each physics simulation sample is computationally expensive. Practitioners are exploring sampling strategies like Latin hypercube and active learning, data representations such as point clouds and graphs, and physics-aware augmentation to build reliable training datasets with limited samples.

I’ve been experimenting with ML models trained on physics simulation output fluid flow and structural fields , and the biggest wall I keep hitting isn’t the model — it’s the data. Generating each simulation sample is computationally expensive, so building a dataset large and diverse enough to train a reliable model is a real challenge. I’d love to hear how others working with scientific or simulation data handle this: Sampling strategy: When each sample is expensive, how are you choosing which cases to generate? Latin hypercube, active learning, adaptive sampling — what’s actually paid off? Data representation: For field data on meshes/grids, are you storing and feeding it as point clouds, voxel grids, graphs, or resampled uniform arrays? What worked for your model type? Augmentation: Are there meaningful augmentation techniques for physics data symmetry, rotation, superposition , or does that risk breaking the underlying physics? Active learning: Has anyone successfully used the model itself to decide which new expensive samples to generate next? Interested in real experience with the “few, expensive samples” regime rather than the usual big-data assumptions.