Ecosystem: High-Throughput Data Generation for ML Surrogates

[jameslehoux]

Labels: ecosystem, machine-learning, phase:4-hpc
Priority: Medium (Strategic for AI citations)

Description

Machine learning researchers are actively trying to train Surrogate AI models to predict tortuosity, but they lack massive datasets of 3D microstructures with accurate, physics-based ground truths.

OpenImpala is perfectly positioned to be the "ground truth generator" for the AI battery community. We should provide an out-of-the-box script/pipeline for high-throughput synthetic data generation.

Acceptance Criteria

• Expand data/create_sample_structure.py to support parameterized generation of stochastic porous media (e.g., overlapping spheres, Gaussian random fields).
• Create an MPI-enabled batch script (examples/generate_ml_dataset.py) that generates $N$ structures, solves the tortuosity for each, and exports a paired dataset (e.g., HDF5 or WebDataset format containing the 3D array and the scalar labels).