Differential equations, meet PyTorch Lightning

What it does

PINA is a Python framework for scientific machine learning built atop PyTorch, PyTorch Lightning, and PyTorch Geometric. It abstracts the usual SciML sprawl into a four-step pipeline: define a Problem, pick a Model, choose a Solver (supervised, PINN, or otherwise), and hand it to the Trainer. The goal is to make experimentation modular—swap a solver or model without rewiring the training infrastructure.

The interesting bit

Instead of treating physics-informed training as a bespoke scripting exercise, PINA inherits PyTorch Lightning’s multi-device scaling and checkpointing, then layers its own Condition and Equation abstractions on top. That means your ODE boundary conditions and neural-operator graph layers share the same training loop mechanics.

Key highlights

• Built on PyTorch Lightning, so multi-GPU training and distributed scaling come built-in rather than bolted-on.
• Modular four-step API (Problem → Model → Solver → Trainer) that tries to keep the experiment structure consistent across data-driven and physics-informed tasks.
• Supports both standard supervised learning and physics-informed networks via the same pipeline; the README shows examples for each.
• Graph neural network support through PyTorch Geometric for problems with geometric structure.
• Published in JOSS, with tutorials and a contributor community.

Verdict

Researchers and engineers who regularly prototype PINNs or neural operators and want Lightning-grade training ergonomics should look here. If you just need a one-off vanilla PyTorch model for a standard benchmark, it’s probably overkill.