PDE-OBS Benchmark Builder

Design reproducible partial-observation PDE benchmarks, generate the exact run code, and report data quality for every selected equation.

Source: private GitHub repository; this public site is a static release mirror.
Main contribution

Build your partial-observation benchmark

Choose PDEs, boundaries, physical regimes, observation masks, and quality gates. The Builder returns reproducible dataset YAML and run code.

Open Benchmark Builder

Read the benchmark contract

Partial observations, controlled

One benchmark workflow from ground truth to quality report

Generate the physical trajectory first, then derive sparse observations as reproducible views. Observation masks never change the underlying ground truth.

1ChoosePDE, boundary, setting, regime, scale
2GenerateDeterministic ground-truth trajectories
3ObserveNine sparse or structured mask views
4ValidateChecksums, provenance, and PDE-specific losses
Full dataset campaign

560,000 / 560,000 samples generated

Open full progress report
3,360complete shards
100%generation complete
7PDE families
0partials or locks
Validation status: every shard passes the unchanged pde_loss ≤ 0.05 gate; the final strict aggregate validator is running separately.
Frozen design space

A benchmark you can inspect before running

Full contract
7PDE families
4boundary protocols
10condition settings
3physical regimes
9observation masks
7PDE loss reports
Scientific status: the code and quality-reporting workflow are available. Paper-grade ground truth still requires the documented numerical-validation gate.
Supporting resource

AI4PDE paper library

The benchmark comes first. The literature library is a supporting index of 301 AI4PDE/AI4SDE papers, with 12 detailed pages. Find work by the physical equation or by the learning method.

Browse by PDEFluids, diffusion, elliptic systems, waves, and more Browse by methodNeural operators, PINNs, diffusion, graphs, and baselines Search all papersCombine PDE, method, venue, year, and keyword filters