Reproduce “Physics-Informed Chebyshev Polynomial Neural Operator for Parametric Partial Differential Equations” on OpenResearch

About this paper

Biao Chen, Jing Wang, Hairun Xie, Qineng Wang, Shuai Zhang, Yifan Xia, Jifa Zhang

The Physics-Informed Chebyshev Polynomial Neural Operator (CPNO) offers a robust, mesh-free deep learning framework for solving parametric Partial Differential Equations by leveraging Chebyshev spectral features and parameter-dependent modulation. This architecture achieves superior accuracy, faster convergence, and enhanced robustness across various PDEs, outperforming existing neural operators and addressing challenges like spectral bias.

Key insights

• Chebyshev polynomials inherently provide numerical stability, with their Gram matrix condition number growing polynomially, significantly outperforming monomial bases which exhibit exponential ill-conditioning.
• The Chebyshev spectral basis allows for a near-optimal and spectrally balanced function space, enabling efficient multi-scale representation and stable gradient flow, thereby mitigating the spectral bias of MLPs.
• The boundedness properties of Chebyshev polynomials and their derivatives ensure numerical stability during both forward and backward propagation, preventing solution divergence and ensuring well-conditioned local optimization sub-problems.

Summary by alphaXiv.