The missing data layer between physics simulations and scientific ML.

PLAID is an open framework for representing, sharing, and learning from datasets of complex physics simulations. It defines a common standard for simulation data and ships a Python library to create, explore, store, and stream them.

Mainstream ML stacks (Hugging Face, PyTorch, TensorFlow) assume data is regular, homogeneous, and columnar. Real simulation data is not: it is hierarchical and multi-zone, with heterogeneous fields, shapes, and metadata, often governed by implicit, solver-specific conventions. Flattening or padding it into tabular form is error-prone, memory-hungry, and erases the physical structure the model should learn from.

• Fidelity — Keep all the complexity of your simulation data — meshes, fields, tags, time, and multiphysics structure — and exploit it directly in ML pipelines.
• Out-of-core datasets — Datasets are accessed sample by sample, so full datasets do not need to be loaded into memory.
• Parallel I/O — save_to_disk can shard sample IDs across multiple processes for fast dataset generation and writing.
• Multiple storage backends — Use CGNS, Hugging Face Datasets, or Zarr through a unified API for local disk, Hub download, and streaming workflows.
• Selective reading — Request only the features you need and, when necessary, only selected indices within large variable arrays.
• Interactive viewer — Launch plaid-viewer to browse local or streamed datasets, inspect samples in 3D, select features, and visualize fields.