System Overview

Problem → Data/Geometry → Model → Physics(PINN) → Solver → Backend → Research

PINNeAPPle is organized as a chain of independent packages, one per stage of this pipeline, connected by plain-data interfaces (ProblemSpec, sampled coordinate tensors, a loss-function callable, a trained model) rather than by inheritance or a shared runtime object. Each stage can be swapped without touching the others: change the preset without touching the network, change the network without touching the physics, change the training policy without touching the physics, change PyTorch for JAX without touching the training policy.

The core packages

Stage	Package
Problem definition	pinneapple_physics.pde_environment
Data container	pinneapple_data
Geometry & sampling	pinneapple_design.geometry
Model architectures	pinneapple_neural.architectures
Physics / residual compiler	pinneapple_physics.pinn_solver, pinneapple_physics.symbolic_pde
Training policy ("Solver")	pinneapple_neural.trainer
Execution backend	pinneapple_tools.compute_backends
Benchmarking / research	pinneapple_tools.benchmark_suite, pinneapple_tools.hpo_experiments

See Package Layers for the full mapping, including the packages that sit around this core pipeline (classical solvers, UQ, transfer/meta-learning, digital twins, design optimization, and more).

A thin top-level convenience layer

The pinneapple top-level package re-exports the pieces most needed to get started (get_preset, build_model, and similar) so a first experiment doesn't require knowing which sub-package each piece lives in. It is a convenience wrapper, not a separate implementation — every function it exposes delegates to one of the packages above.

Compatibility shims

Three packages exist purely to keep older import paths working while the "real" implementation lives elsewhere: pinneapple_models re-exports pinneapple_neural.architectures, pinneapple_solvers re-exports pinneapple_simulation.numerical_solvers, and pinneapple_train re-exports pinneapple_neural.trainer. New code should import from the non-shim location directly.

Why the separation matters

Because each stage only depends on the interface of the stage before it (a ProblemSpec, a batch of coordinates, a model's forward, a loss dict), you can, for example, benchmark the same trained model against a classical numerical solver's reference solution (pinneapple_simulation.numerical_solvers) without either package knowing the other exists — they only share PhysicalSample/tensor data, not code. This is also what makes the higher-layer packages (pinneapple_analysis for UQ/validation/inverse problems, pinneapple_adaptation for transfer/ meta-learning, pinneapple_systems for time series/co-simulation/digital twins, pinneapple_design.design_optimizer for shape optimization) additive rather than invasive: they consume the core pipeline's outputs and add capability on top, without the core pipeline needing to know they exist.