📄 feat(docs/joss_paper) joss paper

.github/workflows/draft-pdf.yml

@@ -0,0 +1,24 @@
+name: Draft JOSS PDF
+on:
+  push:
+    paths:
+      - docs/joss_paper/**
+      - .github/workflows/draft-pdf.yml
+
+jobs:
+  paper:
+    runs-on: ubuntu-latest
+    name: Paper Draft
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Build joss draft PDF
+        uses: openjournals/openjournals-draft-action@master
+        with:
+          journal: joss
+          paper-path: docs/joss_paper/paper.md
+      - name: Upload
+        uses: actions/upload-artifact@v4
+        with:
+          name: paper
+          path: docs/joss_paper/paper.pdf

.gitignore

@@ -4,6 +4,7 @@ tests/post/*.png
 tests/post/*.yaml
 tests/problem_definition/*.csv
 examples/**/*.png
+tests/problem_definition/split.csv
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

docs/joss_paper/paper.bib

@@ -0,0 +1,54 @@
+@article{casenave2025physics,
+  title={Physics-Learning AI Datamodel (PLAID) datasets: a collection of physics simulations for machine learning},
+  author={Casenave, Fabien and Roynard, Xavier and Staber, Brian and Akkari, Nissrine and Piat, William and Bucci, Michele Alessandro and Kabalan, Abbas and Nguyen, Xuan Minh Vuong and Saverio, Luca and Perez, Rapha{\"e}l Carpintero and others},
+  journal={arXiv preprint arXiv:2505.02974},
+  year={2025}
+}
+
+@inproceedings{poinot2018seven,
+  title={Seven keys for practical understanding and use of CGNS},
+  author={Poinot, Marc and Rumsey, Christopher L},
+  booktitle={2018 AIAA Aerospace Sciences Meeting},
+  pages={1503},
+  year={2018}
+}
+
+@article{casenave2024mmgp,
+  title={{MMGP}: a {M}esh {M}orphing {G}aussian {P}rocess-based machine learning method for regression of physical problems under nonparametrized geometrical variability},
+  author={Casenave, Fabien and Staber, Brian and Roynard, Xavier},
+  journal={Advances in Neural Information Processing Systems},
+  volume={36},
+  year={2024}
+}
+@article{kabalan2025elasticity,
+  title={Elasticity-based morphing technique and application to reduced-order modeling},
+  author={Kabalan, Abbas and Casenave, Fabien and Bordeu, Felipe and Ehrlacher, Virginie and Ern, Alexandre},
+  journal={Applied Mathematical Modelling},
+  volume={141},
+  pages={115929},
+  year={2025},
+  publisher={Elsevier}
+}
+
+@article{kabalan2025ommgp,
+  title={{O-MMGP}: {O}ptimal {M}esh {M}orphing {G}aussian {P}rocess Regression for Solving {PDEs} with non-Parametric Geometric Variations},
+  author={Kabalan, Abbas and Casenave, Fabien and Bordeu, Felipe and Ehrlacher, Virginie},
+  journal={arXiv preprint arXiv:2502.11632},
+  year={2025}
+}
+
+@inproceedings{perez2024gaussian,
+  title={{Gaussian process regression with Sliced Wasserstein Weisfeiler-Lehman graph kernels}},
+  author={Carpintero Perez, Rapha{\"e}l and Da Veiga, S{\'e}bastien and Garnier, Josselin and Staber, Brian},
+  booktitle={International Conference on Artificial Intelligence and Statistics},
+  pages={1297--1305},
+  year={2024},
+  organization={PMLR}
+}
+
+@article{perez2024learning,
+  title={{Learning signals defined on graphs with optimal transport and Gaussian process regression}},
+  author={Carpintero Perez, Rapha{\"e}l and Da Veiga, S{\'e}bastien and Garnier, Josselin and Staber, Brian},
+  journal={arXiv preprint arXiv:2410.15721},
+  year={2024}
+}

docs/joss_paper/paper.md

@@ -0,0 +1,59 @@
+---
+title: "PLAID: Physics-Learning AI Datamodel"
+tags:
+- python
+- scientific machine learning
+- data model
+- physics simulation
+date: "07 June 2025"
+
+authors:
+- name: Fabien Casenave
+  orcid: 0000-0002-8810-9128
+  affiliation: 1
+- name: Xavier Roynard
+  orcid: 0000-0001-7840-2120
+  affiliation: 1
+- name: Alexandre Devaux--Rivière
+  orcid: 0009-0001-7474-944X
+  affiliation: 1,2
+affiliations:
+- name: SafranTech, Safran Tech, Digital Sciences & Technologies, 78114 Magny-Les-Hameaux, France
+  index: 1
+- name: EPITA, 14-16 Rue Voltaire, 94270 Le Kremlin-Bicêtre, France
+  index: 2
+bibliography: paper.bib
+---
+
+# Summary
+
+PLAID (Physics-Learning AI Datamodel) is a Python library and data format for representing, storing, and sharing physics simulation datasets for machine learning. Unlike domain-specific formats, PLAID accommodates time-dependent, multi-resolution simulations and heterogeneous meshes. The library provides a high-level API to easily load, inspect, and save data. Beyond basic I/O, PLAID includes utilities for machine-learning workflows. It provides converters to build PLAID datasets from generic tabular data, and a “Hugging Face bridge” to push/pull datasets via the Hugging Face hub. In short, PLAID couples a flexible on-disk standard with a software toolkit to manipulate physics data, addressing the needs of ML researchers in fluid dynamics, structural mechanics, and related fields in a generic fashion. Full documentation, examples and tutorials are available at [plaid-lib.readthedocs.io](https://plaid-lib.readthedocs.io/en/latest/).
+
+
+# Statement of Need
+
+Machine learning for physical systems often suffers from inconsistent data representations across different domains and simulators.  Existing initiatives typically target narrow problems: e.g., separate formats for CFD or for finite-element data, and dedicated scripts to process each new dataset. This fragmentation hinders reproducibility and reuse of high-fidelity data.
+
+PLAID addresses this gap by providing a generic, unified datamodel that can describe many physics simulation data.  It leverages the CGNS standard [@poinot2018seven] to capture complex geometry and time evolution: for example, CGNS supports multi-block topologies and evolving meshes, with a data model that separates abstract topology (element families, etc.) from concrete mesh coordinates.  On top of CGNS, PLAID layers a lightweight organizational structure.
+
+By promoting a common standard, PLAID makes physics data interoperable across projects. It has already been used to package and publish multiple datasets covering structural mechanics and computational fluid dynamics. These PLAID-formatted datasets (hosted on Zenodo and Hugging Face) have supported ML benchmarks, democratizing access to simulation data.
+
+# Functionality
+
+* **Data Model and Formats:** A PLAID dataset is organized within a root folder (or archive), distinctly separating simulation data from machine learning task definitions, as illustrated in \autoref{fig:plaid_dataset_architecture}. The `dataset/` directory contains numbered sample subfolders (`sample_000...`), each holding one or more `.cgns` files under `meshes/` and a `scalars.csv` file. The `dataset/infos.yaml` file contains human-readable descriptions and metadata.  The `problem_definition/` folder provides machine learning context. It includes `problem_infos.yaml` (specifying the ML task inputs/outputs) and `split.csv` (defining train/test splits).  This design supports time evolution and multi-block/multi-geometry problems out of the box.
+
+![Overview of the PLAID dataset architecture.\label{fig:plaid_dataset_architecture}](plaid_architecture.png){ width=80% }
+
+* **Supported Data Types:** PLAID handles scalar, time-series and vector field data on meshes, as well as sample-specific metadata. The `get_mesh(time)` method reconstructs the full CGNS tree for a given timestep, with links resolved if requested (thereby returning the complete mesh). Thus PLAID naturally supports mesh-based simulation outputs with arbitrary element types and remeshing between time steps. Heterogeneity is allowed: missing data is supported, and outputs on testing sets may be missing on purpose to facilitate benchmark initiatives.
+
+* **High-Level API:** The top-level `Dataset` class manages multiple `Sample` objects. Users can create an empty `Dataset()` and add samples via `add_sample()`, or load an existing PLAID data archive by calling `Dataset("path_to_plaid_dataset")`. The `Dataset` object summarizes itself (e.g. printing “Dataset(3 samples, 2 scalars, 5 fields)”) and provides access to samples by ID. Batch operations are supported: one can `dataset.add_samples(...)` to append many samples, or use the classmethods `Dataset.load_from_dir()` and `load_from_file()` to load data from disk, with optional parallel workers. This high-level interface abstracts away low-level I/O, letting users focus on ML pipelines.
+
+* **Utilities:** PLAID includes helper modules for common tasks in data science workflows. The `plaid.utils.split` module provides a `split_dataset` function to partition data into training/validation/testing subsets according to user-defined ratios. The `plaid.utils.interpolation` module implements piecewise linear interpolation routines to resample time series fields or align datasets with different timesteps. The `plaid.utils.stats` module offers an `OnlineStatistics` class to compute running statistics (min, mean, variance, etc.) on arrays, which can be used to analyze dataset distributions. Moreover, a “Hugging Face bridge” (`plaid.bridges.huggingface_bridge`) enables converting PLAID datasets to/from Hugging Face Dataset objects.
+
+# Usage and Applications
+
+PLAID is designed for AI/ML researchers and practitioners working with simulation data. Various datasets, including 2D/3D fluid and structural simulations, are provided in PLAID format in [Hugging Face](https://huggingface.co/PLAID-datasets) and [Zenodo](https://zenodo.org/communities/plaid_datasets). Interactive benchmarks are hosted in a [Hugging Face community](https://huggingface.co/PLAIDcompetitions) on these datasets, providing detailed instructions and PLAID commands for data retrieval and manipulation, see [@casenave2025physics]. These datasets are also used in recent publications to illustrate the performance of the proposed scientific ML methods. In [@casenave2024mmgp; @kabalan2025elasticity; @kabalan2025ommgp],  Gaussian-process regression methods with mesh morphing are applied to these datasets. In [@perez2024gaussian; @perez2024learning] the datasets are leveraged in graph-kernel regression methods applied to fluid/solid mechanics.
+
+In summary, PLAID provides a comprehensive framework for physics-based ML data. By combining a unified data model, support for advanced mesh features, and helpful utilities, it addresses the need for interoperable, high-fidelity simulation datasets. Future enhancements involve developing general-purpose PyTorch dataloaders compatible with PLAID, along with establishing standardized evaluation metrics and unified pipelines for training and inference using the PLAID framework.
+
+# References