PDEBench
PDEBench (2023) is a systematic PDE benchmark platform that provides a unified evaluation environment for data-driven solvers and physics-informed neural networks (PINN). Compared with earlier scattered benchmarks, it is orders of magnitude larger in scale (several TB), diversity (20+ equation types) and task difficulty (low- to high-dimensional), making it the closest thing to a "PDE ImageNet" among public datasets.
Covered equations
- ODE: stochastic differential equations, delayed systems
- PDE: heat, wave, advection-diffusion, Navier-Stokes, shallow water, Burgers
- Multiphysics: heat-fluid coupling, reaction-diffusion
Specs
- Total size: ~5 TB
- Training samples: 10,000+ per equation
- Resolution: 128x128, 256x256, 512x512
- Format: HDF5 (.h5) with initial/boundary conditions and reference solutions
- License: Apache-2.0
Use cases: Training and evaluating surrogate models; ablation studies of PINN / DeepONet / FNO; fair comparison of new architectures; teaching the difference between PDE numerical and data-driven methods.
Note: Download from the official GitHub Release (or S3). The paper should be cited as Takamoto et al., PDEBench, 2023.
- PDEArena · Microsoft Research's open-source PDE benchmark suite with differentiable solvers
- CFDBench · Multiphysics benchmark for computational fluid dynamics covering internal/extern
- TIDE · 2026 3D turbulent DNS benchmark with 5 tasks and controlled generalization split
- DrivAerNet++ · MIT's large-scale automotive aerodynamics dataset: 8,000 parametric car geometri
- The Well · Unified benchmark spanning 16 physical domains, dubbed the 'ImageNet moment of p