ResearchPod Summary
Physics-informed neural networks (PINNs) often struggle with spectral bias and the difficulty of balancing loss terms in strong-form formulations. While Kolmogorov-Arnold Networks (KANs) offer a more flexible, spline-based architecture, they are still typically paired with loss functions that either require high-order derivatives (strong form), are restricted to self-adjoint operators (energy form), or require known fundamental solutions (boundary-integral form).
This paper introduces PG-KINN, which replaces these formulations with a Petrov-Galerkin weak form. In this approach, the KAN acts as the trial space, while an independent, compactly supported, piecewise-polynomial space serves as the test space. By using integration by parts, the authors transfer derivatives onto the test functions, lowering the required differentiation order of the KAN and allowing the solver to handle general, non-self-adjoint, and nonlinear operators without needing fundamental solutions.
PG-KINN demonstrates several key advantages over existing physics-informed methods:
This work provides a bridge between deep learning-based PDE solvers and classical finite element methods. By combining the approximation power of KANs with the mathematical rigor of the Petrov-Galerkin framework, PG-KINN offers a more reliable and versatile tool for computational mechanics. It effectively addresses the limitations of current AI-based solvers in complex engineering scenarios where operator generality and stability are paramount.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.