Skip to content

Benchmark for learning stiff problems using physics-informed machine learning

License

Notifications You must be signed in to change notification settings

DENG-MIT/StiffNet-Benchmark

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 

Repository files navigation

StiffNet-Benchmark

Benchmark problems/datasets for learning stiff multi-physics multi-scale systems using physics-informed machine learning.

Numerical stiffness refers to the challenges in numerically solving systems with widely spaned time/spatial scales. Most complex multi-physics multi-scale systems are numerically stiff, for example, the equations describing the reaction-diffusion in energy conversion, pollution transportation in the atmosphere, and gene/proteins/cell interactions in biological systems. While physics-informed machine learning, such as Physics-Informed Neural Network, Neural Ordinary Differential Equations, Deep Operator Learning, have shown success in many non-stiff systems, it has been shown the stiffness could fail many physics-informed machine learning algorithms. However, it is usually a painful process to prepare the datasets for stiff problems using classical numerical methods due to the nature of multi-physics, and each problem requires specific domain knowledge. This repository aims to ease the data/model preparation stage by formulating a collection of benchmark tasks for testing the performance of new algorithms on physics-informed machine learning in stiff systems.

RoadMap

Candidate Problems

Problems refer to experimental/numerical datasets for stiff problems.

For numerical problems, problems will appear as codes relying on open-source packages with detailed instructions for generating dataset.

Ordinary Differential Equations

  • ROBER (Python/Julia): A classical simple yet challanging benchmarks for stiff solvers.

  • POLLU (Python/Julia): A classical simple yet challanging benchmarks for stiff solvers.

  • Pyrolysis of Propane {Combustion | rely on Cantera} {Coming Soon}: A strong stiff system from the field of combustion.

Partial Differential Equations

  • RUSS (Julia): A classical simple yet challanging benchmarks for stiff solvers.

  • One-dimensional Non-premixed Flame {Combustion | rely on Cantera} {Comming Soon}: A strong stiff system from the field of combustion.

Candidate Machine Learning APIs

Physics-Informed Neural Network (PINN)

E.g., https://github.com/DENG-MIT/Stiff-PINN

Neural Ordinary Differential Equations (NODE)

E.g., https://github.com/DENG-MIT/StiffNeuralODE

(Physics-Informed) DeepONet

(Physics-Informed) Fourier Neural Network

Relevent Links

SciML Benchmarks

About

Benchmark for learning stiff problems using physics-informed machine learning

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published