Python library supporting Discrete Variational Formulations and training solutions with Collocation-based Robust Variational Physics Informed Neural Networks (DVF-CRVPINN)
New Python library uses discrete weak formulations and robust loss to solve challenging PDEs like Stokes equations.
A team of researchers from AGH University of Science and Technology has developed DVF-CRVPINN, a specialized Python library that introduces a novel approach to solving Partial Differential Equations (PDEs) using discrete weak formulations. The library creates a programming environment for defining discrete computational domains, constructing discrete inner products, and implementing discrete weak formulations with Kronecker delta test functions. This mathematical framework enables neural networks to be trained on solution functions defined over discrete point sets, using discrete finite difference derivatives within automatic differentiation procedures.
The researchers focused on the challenging Stokes equations in two dimensions as their primary computational model, training solutions using discrete weak residuals and the Adamax optimization algorithm. The library's key innovation is its robust loss function that maintains a direct relationship to the true numerical error, providing unprecedented control over accuracy during neural network training. Beyond the Stokes formulation, the team also demonstrated the library's functionality using Laplace problem formulations, proving both the well-posedness and robustness of their approach through rigorous mathematical analysis.
This work represents a significant advancement in Physics-Informed Neural Networks (PINNs), addressing common stability and accuracy issues in traditional implementations. By combining discrete variational formulations with robust loss functions, the DVF-CRVPINN library offers researchers and engineers a more reliable tool for solving complex physical systems where traditional numerical methods might struggle with stability or computational efficiency.
- Library implements discrete weak formulations for PDEs using Kronecker delta test functions and discrete inner products
- Employs robust loss function directly related to true error, providing control over numerical accuracy during training
- Successfully tested on challenging 2D Stokes equations and Laplace problems using Adamax optimization algorithm
Why It Matters
Provides engineers and researchers with more stable, accurate tools for simulating complex physical systems like fluid dynamics and heat transfer.