PINNeAPPle 🍍

Your Physics AI Laboratory — from first principles to real-world systems

Experiment. Learn. Build. Then scale — anywhere.

PINNeAPPle is an open-source Physics AI research and experimentation platform designed to take you from your first physics-informed neural network all the way to robust, production-ready solutions — independent of any specific framework, vendor, or ecosystem.

---

Why PINNeAPPle?

Modern Physics AI ecosystems are powerful — but they assume you already understand:

• How to formulate physical problems correctly
• Which architectures to use (PINNs, operators, surrogates…)
• How to validate physics consistency
• How to benchmark and trust your results

PINNeAPPle is where you build that foundation.

Your physics problem
        ↓
  [ PINNeAPPle ]   ← experiment freely here
    Understand the physics
    Try architectures
    Compare approaches
    Validate results
    Build intuition
        ↓
[ Your Target Stack ]
  (custom infra, HPC, cloud, internal platform, etc.)
  Scale, deploy, integrate

---

Package Structure

PINNeAPPle is organized into 8 mega-modules, each grouping related sub-modules:

pinneapple_physics/
├── pde_environment/    # PDE problem specs, BCs, ICs, presets, RANS
├── pinn_solver/        # PINN compiler, DoMINO domain decomposition
└── symbolic_pde/       # SymPy → autograd residual compiler

pinneapple_neural/
├── architectures/      # SIREN, ModifiedMLP, AFNO, HashGridMLP, MeshGraphNet
├── trainer/            # Trainer, TwoPhase, DDP, Causal, HPC utilities
└── predictor/          # Batched inference, grid evaluation, FlowVisualizer

pinneapple_analysis/
├── uncertainty/        # MC-Dropout, Ensemble UQ, conformal, calibration
├── validation/         # Conservation, BC, symmetry checks vs. reference
└── inverse_problems/   # Noise models, regularizers, EKI, SINDy discovery

pinneapple_adaptation/
├── transfer_learning/  # Fine-tuning, layer freezing, progressive unfreezing
└── meta_learning/      # MAML, Reptile, PDETaskSampler, few-shot adaptation

pinneapple_simulation/
├── numerical_solvers/  # FEM, FDM, FVM, Spectral, SPH, LBM, OpenFOAM, FEniCS
├── particle_dynamics/  # MPM, SPH particles, rigid-body (pure PyTorch)
└── external_solvers/   # OpenFOAM, MATLAB, FMU/Modelica, FEniCS bridges

pinneapple_systems/
├── time_series/        # LSTM, GRU, NBeats, TFT, TCN, XGBoost, HHT, FFT
├── cosimulation/       # Graph co-sim engine: PINNNode, CoSimGraph, CoSimTrainer
└── digital_twin/       # Live twin, sensor streams, EKF/EnKF, anomaly detection

pinneapple_design/
├── geometry/           # SDF library, CSG, physics domains, mesh, NACA airfoil
└── design_optimizer/   # Adjoint, Pareto, Bayesian/evolutionary optimization

pinneapple_tools/
├── visualization/      # CFD-style plots, streamlines, Q-criterion, animations
├── model_export/       # TorchScript, ONNX, CSV, NPZ
├── hpo_experiments/    # Paper discovery, knowledge base, HPO
├── benchmark_suite/    # Arena, leaderboards, transfer/meta benchmark pipelines
└── compute_backends/   # PyTorch (default) + JAX backend abstraction

Additional packages:

• pinneapple_data — UPD dataset
• pinneapple_pdb — physics database
• pinneapple_problemdesign — NLP → PDE agent
• pinneapple_app — FastAPI + frontend web app for benchmarking PINN models on physics problems (Docker-composed backend/frontend)
• pinneapple_arena — YAML/JSON-driven multi-model physics benchmark runner (~80+ architectures, physics losses, UQ, inverse problems)
• pinneapple_blender — export a field/trajectory as a .ply sequence, and optionally build/render a Blender scene via a real local Blender install
• pinneapple_hub — model hub client (push_to_hub/from_pretrained + ModelCard) built on the Hugging Face Hub
• pinneapple_llm — LLM-assisted physics-AI pipeline drafting, gated by a physics-grounded PhysicsGuardrail verification layer
• pinneapple_models — compatibility shim re-exporting pinneapple_neural.architectures (not a separate package)
• pinneapple_perception — extracts physics observations (velocity fields, boundary geometry, modal frequencies) from images, video, and audio
• pinneapple_registry — local, self-hosted artifact registry: versioned model/dataset storage, experiment tracking, and problem-spec history
• pinneapple_solvers — compatibility shim re-exporting pinneapple_simulation.numerical_solvers (not a separate package)
• pinneapple_train — compatibility shim re-exporting pinneapple_neural.trainer (not a separate package)
• pinneapple_worldmodel — generalist Physics Foundation Model trained across many physics domains

---

Installation

pip install pinneapple

With optional extras:

pip install "pinneapple[solvers]"      # numba-accelerated FDM/FEM/LBM
pip install "pinneapple[pinn]"         # SymPy symbolic PDE compiler
pip install "pinneapple[geom]"         # trimesh, meshio, gmsh
pip install "pinneapple[fenics]"       # FEniCS / DOLFINx bridge
pip install "pinneapple[export]"       # ONNX export
pip install "pinneapple[all]"          # everything

---

Three Tiers of Physics AI Experience

Tier 1 — Explorer

"I understand the physics. I want to see what AI can do with it."

from pinneapple_physics import get_preset, solve_pde
from pinneapple_neural import build_model

# Load a 2D Poisson problem preset
spec = get_preset("poisson_2d")

# Build a SIREN network and train it in one call
model = build_model("siren", in_dim=2, out_dim=1, hidden_dim=64, n_layers=4)
result = solve_pde(spec, model, epochs=3000)
result["history"]  # {"loss": [...]}

---

Tier 2 — Experimenter

"I want to test ideas and compare approaches."

from pinneapple_tools.benchmark_suite import Arena

runner  = Arena.from_preset("burgers_1d")
results = runner.compare(["VanillaPINN", "siren"], epochs=2000)
print(results.leaderboard())

---

Tier 3 — Builder

"I want to turn this into a real system."

from pinneapple_neural.trainer import DDPPINNTrainer, DDPTrainerConfig
from pinneapple_tools.model_export import export_onnx
from pinneapple_systems.digital_twin import build_digital_twin, MQTTStream

# Distributed training: one DDPPINNTrainer.setup(rank, world_size) call per
# spawned process, then loss_fn(model, epoch) -> Tensor each step
cfg     = DDPTrainerConfig(backend="nccl", world_size=4)
trainer = DDPPINNTrainer(model, cfg)
trainer.setup(rank=0, world_size=4)
history = trainer.train(loss_fn, n_epochs=10_000)

# Export to ONNX
export_onnx(model, "surrogate.onnx", example_input=x_sample)

# Wrap as a live digital twin
twin = build_digital_twin(model, field_names=["u", "v", "p"])
twin.add_stream(MQTTStream(broker="sensors.local", topic="plant/telemetry", sensor_id="s1", field_names=["u", "v", "p"]))
twin.start()

---

Key Features

Mega-module	Sub-modules	What it does
pinneapple_physics	pde_environment · pinn_solver · symbolic_pde	Define PDEs, compile PINN losses, SymPy → autograd
pinneapple_neural	architectures · trainer · predictor	SIREN/AFNO/MGN models, distributed training, inference
pinneapple_analysis	uncertainty · validation · inverse_problems	UQ, physics consistency checks, parameter inversion
pinneapple_adaptation	transfer_learning · meta_learning	Fine-tune across PDEs, MAML/Reptile few-shot
pinneapple_simulation	numerical_solvers · particle_dynamics · external_solvers	FEM/FDM/SPH/LBM, OpenFOAM/FEniCS bridges
pinneapple_systems	time_series · cosimulation · digital_twin	Forecasting, co-sim graphs, live sensor fusion
pinneapple_design	geometry · design_optimizer	SDF/CSG geometry, adjoint + Bayesian shape opt
pinneapple_tools	visualization · model_export · benchmark_suite · compute_backends	CFD plots, ONNX export, Arena benchmarks, JAX backend

---

Quick Examples

import torch

# ── Physics problem definition ──────────────────────────────────────────────
from pinneapple_physics.pde_environment import get_preset
from pinneapple_physics.pinn_solver import compile_problem

spec   = get_preset("burgers_1d")
losses = compile_problem(spec)

# ── Neural network architectures ────────────────────────────────────────────
from pinneapple_neural.architectures import ModelRegistry, SIREN, AFNO
from pinneapple_neural.trainer import Trainer, TrainConfig
from pinneapple_simulation.numerical_solvers.problem_runner import generate_pinn_dataset

model = ModelRegistry.build("siren", in_dim=2, out_dim=1, hidden_dim=128, n_layers=6)

# One full physics batch (collocation + BC/IC points), re-used every epoch
batch  = generate_pinn_dataset(spec, n_col=4096, n_bc=512)
loader = [{k: (torch.as_tensor(v) if hasattr(v, "dtype") else v) for k, v in batch.items()}]

cfg     = TrainConfig(epochs=5000, device="cuda")
trainer = Trainer(model, losses)
result  = trainer.fit(loader, loader, cfg)

# ── Uncertainty quantification ──────────────────────────────────────────────
from pinneapple_analysis.uncertainty import uq_predict
from pinneapple_analysis.validation import validate_model

uq_result  = uq_predict(model, x_test, method="mc_dropout")
val_report = validate_model(model, spec)

# ── Design optimization ─────────────────────────────────────────────────────
from pinneapple_design.geometry import get_domain
from pinneapple_design.design_optimizer import DesignOptLoop, DesignOptConfig

domain = get_domain("lid_driven_cavity_2d")
x_int  = domain.sample_interior(4096)

# ── Simulation data generation ──────────────────────────────────────────────
from pinneapple_simulation.numerical_solvers import HeatConduction3D
from pinneapple_simulation.numerical_solvers.fdm3d import HeatConfig3D

solver = HeatConduction3D(HeatConfig3D(nx=32, ny=32, nz=32))
data   = solver.solve()

# ── Time series forecasting ─────────────────────────────────────────────────
from pinneapple_systems.time_series import NaiveForecaster

forecaster = NaiveForecaster()
forecaster.fit(train_series)
forecast = forecaster.predict(24)

# ── Benchmarking ────────────────────────────────────────────────────────────
from pinneapple_tools.benchmark_suite import Arena

runner = Arena.from_preset("poisson_2d")
result = runner.run("siren", epochs=5000)
print(result.summary())

---

Examples

Folder	What it covers
examples/pde_environment/	PDE presets, BCs, problem specs
examples/pinn_solver/	PINN compiler, symbolic losses
examples/architectures/	Model registry, SIREN, AFNO, GNN, operators
examples/trainer/	Training loops, DDP, HPC, AMP
examples/numerical_solvers/	FEM, FDM, FVM, SPH, LBM, spectral
examples/time_series/	Forecasting, backtesting, uncertainty
examples/geometry/	SDF, CSG, mesh, airfoil generation
examples/benchmark_suite/	Arena, YAML configs, leaderboards
examples/hpo_experiments/	Paper discovery, knowledge base
examples/data_pipeline/	UPD datasets, Zarr, active learning
examples/physics_db/	Physics database, NASA/Earthdata
examples/problem_designer/	NLP → PDE agent

---

Philosophy

If you can't validate it, you shouldn't deploy it.

Physics AI is about:

• Correct formulations
• Reliable validation
• Understanding failure modes
• Making informed decisions

---

Positioning

	PINNeAPPle
Vendor lock-in	❌ Not tied to any vendor
Just a PINN library	❌ Much more than that
Just experimentation	❌ Bridges to production
✅ What it is	A controlled environment to design, test, and validate Physics AI systems

---

Citation

If you use PINNeAPPle in academic research, technical reports, benchmarks, or industrial publications, please cite the framework.

BibTeX

@software{pinneapple2026,
  title        = {PINNeAPPle: An Open-Source Physics AI Research and Experimentation Platform},
  author       = {Barros, Yan and Contributors},
  year         = {2026},
  url          = {https://github.com/PINNeAPPle-Labs/PINNeAPPle},
  version      = {0.6.1}
}

---

Support the Project

If this project makes sense to you, give it a star ⭐

It helps grow the ecosystem, attract contributors, and build a real standard.

---

Built for researchers and engineers who take physics seriously.