STAR-PINN v1.0.0: Stacked Adaptive Residual Physics-Informed Neural Networks
Published:
STAR-PINN: Stacked Adaptive Residual Physics-Informed Neural Networks. [GitHub]
We are pleased to announce the public release of STAR-PINN v1.0.0, an open-source implementation of the Stacked Adaptive Residual Physics-Informed Neural Network framework for solving magnetic diffusion problems.
Core Concept
Instead of relying on a single monolithic PINN, STAR-PINN progressively refines the solution through a sequence of residual-learning PINNs:
- PINN₀ learns the initial physical solution
- Additional PINNs learn residual corrections
- Adaptive mixing progressively refines predictions
- Each stage improves accuracy, stability, and physical consistency
This approach treats learning as an iterative physics-aware refinement process, enabling more accurate solutions for challenging electromagnetic diffusion problems.
Related Publication
This repository accompanies our paper:
Dodge et al., “STAR-PINN: Stacked Adaptive Residual Physics-Informed Neural Networks for Magnetic Diffusion Problems”, IEEE Access, 2026. [Link]
Repository Features
- Physics-informed deep learning framework
- Residual-based stacked PINN architecture
- Adaptive network weighting strategy
- Magnetic diffusion benchmarks
- Reproducible training and evaluation scripts
- Open-source implementation for research and education
We hope this repository will be useful for researchers and engineers working in Scientific Machine Learning, Computational Physics, and Computational Electromagnetics.
Contributions, suggestions, and feedback are always welcome.