Sinc Kolmogorov-Arnold network and its application for solving PDEs with singularities

Researchers propose SincKANs, a neural network architecture combining Sinc interpolation with Kolmogorov-Arnold Networks to improve function approximation and solve partial differential equations. The approach demonstrates superior performance compared to existing methods, particularly for functions with singularities, offering potential advances in physics-informed machine learning.

This research represents a technical advancement in neural network design that bridges numerical analysis and deep learning. Sinc interpolation, a well-established method in numerical analysis, is being integrated into Kolmogorov-Arnold Networks (KANs), a relatively recent innovation that uses learnable activation functions instead of fixed ones like ReLU. The combination addresses a critical limitation in neural networks: handling functions with discontinuities or singularities, which are common in real-world physics simulations.

The broader context involves the growing interest in KANs as alternatives to traditional multilayer perceptrons. Since their introduction, researchers have experimented with various function representations within the KAN framework. This work systematically demonstrates that Sinc interpolation outperforms previous approaches across multiple test cases. The significance lies in improved accuracy for physics-informed neural networks (PINNs), which solve partial differential equations by training neural networks to satisfy both differential equations and boundary conditions.

For the AI and scientific computing communities, this development has practical implications. Researchers working on fluid dynamics simulations, quantum mechanics problems, and other domains involving singularities could benefit from more accurate neural network approximations. Better PINNs reduce computational costs compared to traditional numerical solvers while maintaining accuracy. This could accelerate adoption of neural networks in scientific computing pipelines.

Looking forward, the key question is whether SincKANs will gain adoption in production scientific computing workflows. Additional research on computational efficiency, scalability to higher-dimensional problems, and comparison with other recent KAN variants will determine practical impact. The work establishes a foundation that other researchers can build upon to further optimize neural network architectures for domain-specific problems.

• →SincKANs combine Sinc interpolation with Kolmogorov-Arnold Networks for improved function approximation, especially with singularities.
• →Physics-informed neural networks demonstrate better performance using Sinc-based activation functions across tested examples.
• →The approach leverages established numerical analysis techniques to address known limitations in neural network design.
• →Implementation could reduce computational costs in scientific computing by improving PINN accuracy and convergence.
• →Further research needed on scalability and real-world performance in higher-dimensional scientific problems.