PLAID
Physics Learning AI Data Model — turning complex physics simulations into AI-ready data.
Overview¶
PLAID (Physics Learning AI Data Model) turns complex physics simulations into AI-ready datasets. It preserves the full structure of each simulation — meshes (including remeshing between time steps), physical fields, mesh tags, temporal evolution, multiphysics couplings, and associated metadata — and exposes it through a unified API designed for storing, streaming, visualizing, and learning from massive, heterogeneous datasets.
Key features¶
-
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_diskcan 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-viewerto browse local or streamed datasets, inspect samples in 3D, select features, and visualize fields.
Get started¶
-
Quickstart
Install PLAID and run your first example in minutes.
-
Concepts
Learn the core abstractions: samples, datasets, and problem definitions.
-
Examples & tutorials
Walk through end-to-end examples and notebooks.
-
API reference
Browse the complete Python API.
About¶
PLAID has been developed at SafranTech, the research center of the Safran group. The source code is hosted on GitHub — contributions, issues, and feedback are welcome.