TL;DR
The Physics-Learning AI Datamodel (PLAID) introduces a unified, extensible open-source standard and Python library built on CGNS for physics simulation datasets. Alongside the data format, the authors release six diverse simulation benchmarks spanning structural mechanics and computational fluid dynamics (CFD), evaluate several machine learning surrogate architectures, and provide an open continuous benchmarking platform on Hugging Face.
What Problem Does This Paper Address?
While machine learning surrogates promise orders-of-magnitude speedups over classical numerical solvers (such as finite element or finite volume methods), progress in scientific machine learning (SciML) remains heavily fragmented. Unlike natural language processing or computer vision, which thrive on massive, standardized corpora:
- Physics datasets often rely on rigid, ad-hoc file formats.
- Most benchmarks are restricted to simple regular Cartesian grids or narrow physical regimes.
- Realistic engineering setups—involving non-linear constitutive laws, dynamic remeshing, mixed-element unstructured meshes, variable topologies, and boundary condition tags—are severely underrepresented.
PLAID addresses this gap by defining a standardized datamodel that natively supports high-dimensional, unstructured, multi-physics simulations.
Methods
- The PLAID Standard & Library: Built on the established CFD General Notation System (CGNS), PLAID provides a unified hierarchical structure stored as human-readable YAML/CSV files paired with binary CGNS data (viewable in ParaView) or hosted via Hugging Face for streaming and caching. It natively handles time series, multi-zone meshes, and varying node/element tags.
- Evaluated Surrogate Architectures:
- MeshGraphNets (MGN): Encodes meshes as graphs and performs message passing across spatial neighbors.
- Mesh Morphing Gaussian Processes (MMGP): Maps non-parameterized geometries onto a reference geometry via morphing and finite-element interpolation, reduces dimensionality using PCA, and performs Gaussian process regression.
- Vi-Transformer & Augur: Partition large unstructured meshes into local point-cloud patches using Metis decomposition, pass patch tokens through a Transformer, and decode back to node-wise fields. Augur incorporates a global context tensor to prevent boundary discontinuities across patches.
- Domain Agnostic Fourier Neural Operators (DAFNO): Uses FFTs within layers alongside smoothed domain masks to capture dynamics across irregular geometries.
- Modulated Aerodynamic Resolution-Invariant Operator (MARIO): Uses coordinate-based implicit neural representations (INRs) with Fourier feature embeddings, conditioned via hypernetworks on learned shape representations from signed distance functions (SDF).
Datasets
PLAID introduces six new datasets and three standardized variants of an existing benchmark:
- Tensile2d: 2D quasistatic non-linear structural mechanics (elastoplasticity with large strain variations) solved with Z-set.
- 2D_MultiScHypEl: 2D multiscale hyperelasticity with varying topology (voids/inclusions) solved with DOLFINx.
- 2D_ElPlDynamics: 2D dynamic elasto-plastic deformation featuring element erosion and failure solved with OpenRadioss.
- Rotor37: 3D steady-state compressible Reynolds-Averaged Navier-Stokes (RANS) flow around a transonic compressor blade solved with elsA.
- 2D_profile: 2D supersonic compressible RANS simulations around morphing aerodynamic profiles solved with elsA.
- VKI-LS59: 2D transonic/supersonic compressible RANS turbine blade cascade flow exhibiting shock waves, solved with BROADCAST.
- AirfRANS (Original, Clipped, Remeshed): External subsonic RANS aerodynamic flow around airfoils formatted under PLAID.
Key Insights and Ideas
- Message passing hits a distance wall: Standard GNNs like MGN struggle with long-range dependencies unless networks become prohibitively deep, making them less competitive on large, global fluid-structure interactions.
- Transformers scale via graph partitioning: By clustering mesh subdomains into tokens (using Metis), Transformers achieve global context awareness without quadratic $O(N^2)$ node-level memory bottlenecks. Aggregating a global latent vector (as in Augur) prevents visual and numerical seams at patch boundaries.
- Mesh morphing unlocks tabular ML: When geometry changes smoothly without topological changes, morphing meshes onto a reference shape allows PCA and Gaussian process regression (MMGP) to drastically outperform complex deep learning baselines in accuracy and training speed.
- Specialized handling for shocks: In datasets with sharp discontinuities (such as shock waves in VKI-LS59), linear PCA decoders struggle, requiring non-linear polynomial decoders or coordinate-based implicit neural representations (MARIO).
Key Results
Evaluated using Relative Root Mean Square Error (RRMSE) averaged across predicted fields and scalar outputs:
- Tensile2d: MMGP led with a total error of 0.0027, outperforming Augur (0.0154) and MGN (0.0673).
- 2D_MultiScHypEl: Augur performed best with 0.0232 total error, compared to MGN's 0.0437 (MMGP could not handle the topological variations of inclusion holes).
- 2D_ElPlDynamics: Evaluated with DAFNO autoregressively over time, reaching 0.0158 total error.
- Rotor37: MMGP yielded the lowest error (0.0019), followed by Augur (0.0033) and MGN (0.0074).
- 2D_profile: MMGP achieved 0.0480 total error versus MGN (0.0593) and Vi-Transformer (0.0638).
- VKI-LS59: MARIO obtained the best total error (0.0124), outperforming MMGP (0.0312), Augur (0.0362), and MGN (0.0684).
Limitations and Future Work
- Model-Specific Constraints: MMGP currently cannot accommodate changing mesh topology (e.g., holes appearing or disappearing). DAFNO requires re-interpolating between irregular simulation meshes and regular Cartesian grids. MARIO is computationally heavy, taking ~30 hours on an A100 GPU for a single run.
- Roadmap: The authors plan to implement generic PyTorch data loaders directly into the PLAID library, standardize training and evaluation pipelines, and expand the repository with larger 3D industrial simulation suites. Community submissions are continuously accepted via Hugging Face competitions.
Summary updated