Aeroscience-Computations-Analysis-Lab

Aeroscience-Computations-Analysis-Lab / underPINN

Public

A modular, JAX-based framework for Physics-Informed Neural Networks, designed for scalable PDE-constrained learning with support for domain decomposition, attention mechanisms, and performance-oriented training.

aeroscience-computations-analysis-lab.github.iounderPINN deep-learning jax neural-network physics-informed-neural-networks pinn
19
5
100% credibility
Found May 31, 2026 at 19 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

underPINN is a research framework that uses neural networks to solve partial differential equations and ordinary differential equations from physics and engineering, featuring GPU acceleration, adaptive training techniques, and pre-built examples for fluid dynamics, heat transfer, and wave propagation problems.

How It Works

1
🔬 You discover a physics-solving tool

A researcher or student learns about underPINN — a framework that uses AI to solve complex physics equations like fluid flow, heat transfer, and wave propagation.

2
📦 You install the framework

You download and install the tool on your computer, which sets up everything needed to start running physics simulations.

3
âš¡ You run your first simulation

Using a simple command, you launch a pre-built example like the 1D wave equation or heat diffusion problem to see the framework in action.

4
🧠 Your neural network learns physics

The AI model trains by itself, discovering the mathematical rules of your physics problem — watching it figure out waves bouncing or heat spreading feels almost magical.

5
You choose your approach
📋
Use a ready example

Pick from Burgers equation, lid-driven cavity flow, airfoil aerodynamics, or pipe flow — the settings are already configured for you.

🔧
Customize your problem

Adjust the geometry, material properties, and boundary conditions to match your exact research or engineering problem.

6
📊 You see your results

Beautiful visualizations appear showing your solution — contour plots of temperature, pressure fields, or velocity maps of fluid flow.

✅ You solved your physics problem

Your neural network found an accurate solution to your PDE, with error metrics and convergence plots confirming the quality of your results.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 19 to 19 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is underPINN?

underPINN is a Python framework for solving partial differential equations using physics-informed neural networks. It combines classical collocation-based PINNs with finite basis domain decomposition, attention mechanisms, and adaptive training strategies, all compiled to GPU via JAX and XLA. You define your PDE, pick a network architecture, and train with automatic loss balancing, checkpointing, and hyperparameter sweeps through a CLI that requires zero Python code.

Why is it gaining traction?

The framework ships with a built-in library of PDEs ranging from 1D Burgers to 3D pipe flow and compressible Euler, so you can benchmark against known problems immediately. Its restart system lets you interrupt long training runs and resume from the last snapshot, while the `lax.scan` acceleration fuses hundreds of gradient steps into a single compiled kernel for 50-500x less Python dispatch overhead on GPU. Transfer learning support lets you warm-start new parameter regimes from trained checkpoints, and inverse problems let you recover unknown physics constants from sparse noisy data.

Who should use this?

Computational physicists and aerospace engineers who need GPU-accelerated PDE solving without building infrastructure from scratch. Researchers benchmarking PINN architectures against canonical fluid dynamics and heat transfer problems will appreciate the built-in evaluators and reporting suite. Teams running parametric studies will value the YAML-driven sweep CLI for Cartesian hyperparameter searches.

Verdict

With only 19 stars and a 1.0% credibility score, underPINN is clearly early-stage and lacks the community backing of established projects, but the feature set is unusually complete for a research framework at this maturity level. The documentation is thorough and the CLI-first design makes it accessible without reading source code, making it worth evaluating for physics-informed machine learning work where GPU performance and reproducibility tooling matter.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.