i207M/PINNacle

[NeurIPS 2024] Codebase for PINNacle: A Comprehensive Benchmark of Physics-Informed Neural Networks for Solving PDEs.

What it solves

PINNacle provides a standardized way to evaluate and compare different Physics-Informed Neural Network (PINN) variants. It addresses the lack of a comprehensive benchmark for solving partial differential equations (PDEs) by implementing multiple PINN methodologies and providing a challenging dataset for performance comparison.

How it works

The project implements a wide range of PINN variants, categorized by their approach to improving PDE solving:

  • Loss reweighting: Methods like PINN-LRA and PINN-NTK.
  • Collocation points resampling: The RAR method.
  • Optimizers: A specialized optimizer called MultiAdam.
  • Loss functions: Regularization terms (gPINN) and variational formulations (hp-VPINN).
  • Architecture changes: Adaptive activation functions (LAAF, GAAF) and domain decomposition (FBPINN).

Who it’s for

Researchers and engineers working with physics-informed machine learning to solve differential equations and optimize the performance of neural networks in scientific computing.

Highlights

  • Implements 11 different PINN variants including vanilla PINNs and advanced architectural and optimization strategies.
  • Includes a new challenging dataset specifically designed for PDE solving benchmarks.
  • Supports a unified benchmark script to run multiple test cases with customizable settings.

Related

  • Project
    SciML/NeuralPDE.jl

    A Julia package for solving partial differential equations using physics-informed neural networks (PINNs) and neural stochastic differential equations.

  • Project
    PINA-org/PINA

    A unified Python framework for Scientific Machine Learning that simplifies the creation of PINNs, Neural Operators, and data-driven models.

  • Project
    lululxvi/deepxde

    A scientific machine learning library for solving differential equations and learning operators using physics-informed neural networks and other advanced AI architectures.

  • Project
    Xilinx/finn

    An experimental framework from AMD Research that enables fast, scalable inference of quantized neural networks on FPGAs by generating customized dataflow architectures.

  • Dispatch
    Google DeepMind Discovers New Fluid Dynamics Singularities Using AI

    Google DeepMind has developed a novel AI method using Physics-Informed Neural Networks (PINNs) to discover new families of unstable singularities in fluid dynamics equations, providing a path toward solving long-standing mathematical challenges.