The original PINNs paper code: a museum piece with a warning label

What it does

Trains neural networks to solve and discover nonlinear partial differential equations by baking physical laws directly into the loss function. Handles both “forward” problems (inferring solutions) and “inverse” problems (discovering the equations themselves) using continuous or discrete time formulations. The resulting models are fully differentiable surrogates that supposedly need less data because physics does the regularizing.

The interesting bit

This is the 2017-2019 reference implementation that coined the term “PINNs” and spawned an entire subfield. The authors split the work into two neat acts: Part I solves equations, Part II discovers them — a rare case where the academic paper structure maps cleanly onto actual code organization.

Key highlights

• Original implementation of the Raissi et al. framework from the Journal of Computational Physics
• Covers both data-driven solution and data-driven discovery of PDEs
• Supports continuous-time and discrete-time algorithm variants
• ~5,900 stars suggests it was the gateway drug for many researchers
• Authors explicitly recommend PyTorch, JAX, and TensorFlow v2 successors

Caveats

• No longer maintained; authors themselves point to modern reimplementations
• No candidate images provided, so visual validation of results is absent from the README
• The “data-efficient” claim is stated but not quantified in the README

Verdict

Worth a skim for historical context or citation archaeology, but start with the PyTorch/JAX/TF2 forks if you actually want to train something. Skip entirely if you’re looking for production-ready tooling — this is a paper artifact, not a framework.