Physics-Informed Neural Networks (PINNs)

Report on Current Developments in Physics-Informed Neural Networks (PINNs)

General Direction of the Field

The field of Physics-Informed Neural Networks (PINNs) is experiencing a surge in innovation, with recent developments focusing on enhancing the robustness, accuracy, and efficiency of these models. Researchers are increasingly exploring hybrid approaches that combine traditional numerical methods with modern machine learning techniques to address the limitations of existing PINNs. Key areas of advancement include the integration of domain decomposition, adaptive basis functions, and hybrid neural architectures to better capture the complexities of partial differential equations (PDEs). Additionally, there is a growing emphasis on data-efficient training methods and the development of neural solvers that can handle parametric variations in PDEs. These innovations are not only improving the performance of PINNs but also broadening their applicability to a wider range of physical problems, particularly in fluid mechanics and multiphysics simulations.

Noteworthy Innovations

  1. Fourier PINNs: A novel approach that augments PINNs with pre-specified, dense Fourier bases, enabling better learning of high-frequency components without restrictions on boundary conditions or problem domains.

  2. HyResPINNs: Introduces adaptive hybrid residual networks that combine neural and radial basis function components, significantly enhancing accuracy and robustness in solving challenging PDEs.

  3. Transport-Embedded Neural Architecture: Redefines physics-aware neural models in fluid mechanics by following the transport equation, successfully capturing temporal changes in high Reynolds number flows and preventing false minima.

  4. Learning a Neural Solver for Parametric PDE: Proposes a data-driven approach to solving parametric PDEs, integrating physical loss gradients with PDE parameters to enhance the stability and convergence of physics-informed methods.

These advancements are poised to drive the next wave of progress in PINNs, making them more versatile and effective tools for solving complex physical problems.

Sources

Towards Model Discovery Using Domain Decomposition and PINNs

Fourier PINNs: From Strong Boundary Conditions to Adaptive Fourier Bases

HyResPINNs: Adaptive Hybrid Residual Networks for Learning Optimal Combinations of Neural and RBF Components for Physics-Informed Modeling

FastLRNR and Sparse Physics Informed Backpropagation

Transport-Embedded Neural Architecture: Redefining the Landscape of physics aware neural models in fluid mechanics

MelissaDL x Breed: Towards Data-Efficient On-line Supervised Training of Multi-parametric Surrogates with Active Learning

Learning a Neural Solver for Parametric PDE to Enhance Physics-Informed Methods

Built with on top of