Skip to content

How PDEForge Compares

An honest feature comparison with the main alternatives, as of mid-2026. Corrections welcome — open an issue.

PDEForge APEBench/exponax PDEBench The Well PDEArena py-pde
Delivery generate on demand generate on demand download (+per-PDE scripts) download (15 TB) download DIY solver
One-call multi-physics API yes yes (scenarios) no no no no
Any resolution yes yes regeneration is DIY no no yes (DIY)
Any parameters yes yes DIY no partial yes (DIY)
Train/val/calibration/test yes — native no no no (3-way) no (3-way) no
OOD splits by parameter range yes no no no no no
Multi-fidelity pairs yes no no no no no
FEM / complex geometry yes (FEniCSx) no (periodic only) no fixed datasets no no
Stochastic PDEs yes no no some (fixed) no yes
Framework-agnostic output NumPy/HDF5/zarr JAX arrays HDF5 HDF5/torch HDF5 NumPy
GPU acceleration optional JAX backend JAX-native n/a n/a n/a numba
Base install NumPy/SciPy (pip) JAX (pip) heavy loader only medium pip
Convergence-verified data yes (pdeforge.verify) no no no no no
PDE count 30+ (and growing) ~46 11 families 16 datasets 5 families user-defined

When to use something else. Pretraining a foundation model on tens of terabytes of heterogeneous physics: The Well. Benchmarking autoregressive emulator rollouts with differentiable-solver training in JAX: APEBench. Comparing against the literature's fixed reference numbers: PDEBench's published datasets. A PDE we don't implement and you want to hand-discretise: py-pde, Dedalus, or PhiFlow.

When to use PDEForge. Controlled studies that need data at YOUR resolution and YOUR parameters; anything involving calibrated uncertainty (conformal prediction needs the calibration split we ship natively); distribution-shift and multi-fidelity experiments; complex-geometry flow data without writing FEM code; reproducibility — every dataset regenerates from its own metadata (pdeforge.reproduce).