PINNeAPPle Labs

Physics AI with evidence, not vibes.

We build Physics-Informed Neural Networks and verification tooling that show their work — real execution, real residuals, real evidence, never a confidence number pulled from nowhere.

Live demos

Four working systems, not slide decks. Click through and run them yourself.

Verification

VeriPhysics

Submit a physics problem, watch it execute, and get a per-check evidence breakdown — guardrail, convergence, and more — with the real number and source behind each one, plus a downloadable PDF report.

Run VeriPhysics →
Simulation Surrogate

Surrogate Speed

A trained neural surrogate predicting heatsink temperature in microseconds, checked every time against a real independent 3D finite-element solve.

Run the surrogate →
Digital Twin

Predictive Twin

A real-time river digital twin, calibrated live on a real US river gauge's own recent discharge data, forecasting behavior and checked against live data as it arrives.

Watch the live twin →
Get involved

Real Physics Challenge

Bring a real engineering problem. We test it against Physics AI, for free, in exchange for feedback — no sales pitch, one real experiment.

Bring your problem →

Real results

Every number below came from an actual run, not an estimate.

176,267×
faster than a real independent 3D FEA solve, averaged across 7 curated heatsink designs — while staying within 4.46% average / 7.35% worst-case agreement with it.
Surrogate Speed demo
Live
digital twin calibrated on real-time public USGS river gauge data, continuously checked against the live data still arriving — not a synthetic dataset.
Predictive Twin demo
Per-check
evidence, not one opaque score: every VeriPhysics run shows the real residual, convergence, and calibration numbers behind its trust score, with "not run" shown honestly when a check didn't execute.
VeriPhysics demo
Caught it
Ran VeriPhysics's guardrail against an already-trained PINN from a real, confidential industrial engagement — a physical wear model built for a real client. The model's own reported training loss looked healthy, but the guardrail's per-check breakdown showed its physics residual was hundreds of times over a trustworthy threshold: the network had learned to hit its final target while quietly violating the governing equation along the way. A single aggregate loss number hid exactly the kind of failure this tool exists to catch.
Internal verification run, real client model — details kept confidential

Open source

The foundation is public, free, and on PyPI — real repositories, real releases, not vaporware.

PINNeAPPle

The core engine — PINN/FNO/DeepONet/GNN architectures, solvers, and digital twin tooling. Everything else here is built on top of it.

★ 9 on GitHub · PyPI 0.5.0

pinnfactory

A lightweight, standalone framework for building PINNs from symbolic PDE definitions (SymPy) with automatic differentiation in PyTorch — for anyone who wants PINNs without the whole platform.

★ 71 on GitHub · PyPI 0.1.0

pinnaitor

A general-purpose agentic framework (ELM AI Agentic Framework) — an earlier-stage project, not yet wired into the rest of the PINNeAPPle story.

★ 22 on GitHub

This is the honest split: the 3 repos above are genuinely open source. The demos in "Live demos" (VeriPhysics, the surrogate, the digital twin) run on top of this engine but are commercial products, not open-source releases.

About

Yan Barros

Physics AI Lead Engineer & SciML Researcher

Founder of PINNeAPPle Labs, building physics-informed machine learning that has to show its work — execution, verification, and evidence as the product, not an afterthought.

Connect on LinkedIn →