🎉 "A Pseudo-Time Stepping and Parameterized Physics-Informed Neural Network Framework for Navier-Stokes Equations"
This poster introduces P2PINN: A Pseudo-Time Stepping and Parameterized Physics-Informed Neural Network Framework for Navier-Stokes Equations. Our work addresses two major limitations of traditional Physics-Informed Neural Networks (PINNs): unstable training and the necessity for retraining when PDE parameters change.
P2PINN tackles these issues by:
- Integrating a pseudo-time stepping method to significantly stabilize training, even for complex problems like the Navier-Stokes equations.
- Employing a dedicated encoder network that extracts hidden representations of PDE parameters. This enables the network to learn a generalized solution across a wide parameter range by training on just two PDE parameters.
The key benefits of P2PINN are enhanced training stability and remarkable computational efficiency. Our framework achieves a 2–4 orders of magnitude speedup compared to standard PINNs, all while maintaining high accuracy. This allows for rapid interpolation and extrapolation to unseen parameter sets.
This research is a collaborative effort with Dr. Xiong Xiong from Northwestern Polytechnical University. We wish him the very best for a successful and early graduation!
Article Website
You can find the article published here:
