PINNeAPPle-Labs/PINNeAPPle
The core: PINN, FNO, DeepONet, GNN, solvers, geometry, UQ, digital twins and REST serving.
The library
One pipeline, from the written PDE to the served prediction.
Every layer is an independent Python package on PyPI — use only the part you need. No layer hides the one below it: you see the mesh, the residual, the architecture and the error.
Architecture
Layered architecture
ApplicationAPPS & SERVING
pinneapple_apppinneapple_labpinneapple_hubpinneapple_orchestrationpinneapple_llmREST / FastAPI
ModelsNEURAL
pinneapple_neuralpinneapple_modelsPINNFNODeepONetGNNLSTMVPINN
Physics & solversPHYSICS
pinneapple_physicspinneapple_solverspinneapple_simulationpinneapple_pdbOpenFOAMFEniCS/dolfinxCalculiX
AnalysisANALYSIS & UQ
pinneapple_analysispinneapple_veriphysicspinneapple_decisionpinneapple_adaptationUQEnKFMonte Carlo
Data & geometryDATA
pinneapple_datapinneapple_designpinneapple_catalogpinneapple_pdbpinneapple_twin3dpinneapple_blender
CoreCORE
pinneapple_corepinneapple_problemdesignpinneapple_registrypinneapple_toolspinneapple_securitypinneapple_systems
Why this matters: the same library that trains the network also runs the reference solver. Comparing a model against itself is the most common mistake in Physics AI — here the reference is always external to training.
inverse_heat.py
# the h nobody measures, from a handful of thermocouples from pinneapple_physics import HeatConduction1D from pinneapple_neural import PINNTrainer problem = HeatConduction1D(conductivity=50.0, length=0.30, learn_h=True) model = PINNTrainer(problem, layers=[3, 64, 64, 1]) report = model.fit(thermocouples=readings, epochs=5000) print(report.h, report.h_std) # 14.8 0.3 (true value: 15) report.verify_against("finite-volume") # per-field error, error map
Get started
The equation goes in as text
One-command install, no build. The same package runs the problem, the training, the comparison against the solver and the report.
What you get on day one
- A reproducible environment — 248 examples with versioned output, to compare against your own result.
- Public benchmarks — SSQN Helmholtz, LBM vortex shedding, LPM design space, latency and I/O.
- Real interoperability — OpenFOAM, CalculiX, FEniCS/dolfinx, Gmsh, VTK, STL, USD and glTF in one session.
- Industry standards — AAS, DTDL, NGSI-LD, Eclipse Ditto and OPC UA for what the plant already consumes.
Open source
The base is public. So is verification.
Apache-2.0, on PyPI, with versioned releases. What is open is genuinely open — and what is a product is called a product.
PINNeAPPle-Labs/pinnfactory
A lightweight, standalone framework for PINNs from PDEs in SymPy, with PyTorch autodiff.
PINNeAPPle-Labs/PINNeAPPle-Benchmark
Public, reproducible benchmarks: SSQN Helmholtz, LBM vortex shedding, LPM, latency, shard balance.
PINNeAPPle-Labs/pinnaitor
A general-purpose agentic framework (ELM AI Agentic Framework), at an earlier stage — said plainly.
The honest split: The four repositories above are genuinely open source. The products are delivered as engineering, not as an open-source release.
Bring your case
Got a simulation that is slow, expensive or hard to verify?
Send it over. We come back with a closed scope and a measurable success criterion — or we tell you the classical method does it better.