SirenFNO: Efficient and Full Frequency Learning of Fourier Neural Operators

Researchers introduce SirenFNO, a neural network framework that improves Fourier Neural Operators by eliminating frequency truncation limitations and enabling full-spectrum learning. The approach achieves 4-15x parameter reduction while maintaining discretization invariance, with functional decomposition variants reaching up to 73x fewer parameters across multiple PDE benchmarks.

SirenFNO addresses a fundamental constraint in current machine learning approaches to solving partial differential equations. Traditional Fourier Neural Operators rely on frequency truncation—limiting the mathematical modes they can learn—to maintain computational efficiency. This architectural choice creates spectral bias that systematically underperforms on PDEs with high-frequency oscillations, a limitation impacting physics simulations, engineering design, and scientific computing applications.

The breakthrough combines sinusoidal representation networks with mode-wise kernel parameterization to learn complete frequency spectra without truncation. By maintaining constant parameter counts independent of discretization resolution, SirenFNO removes a key bottleneck while substantially reducing model size. The addition of functional tensor decomposition techniques further compresses parameters while improving learning efficiency, suggesting the approach scales well across problem domains.

For the broader AI infrastructure sector, this development matters because efficient surrogate models for PDEs underpin scientific machine learning applications worth billions in potential value—from climate modeling to drug discovery to materials science. Reducing parameter counts by 10-70x while improving accuracy translates directly to lower computational costs and faster inference times, making neural operator approaches more viable for production deployment. Organizations developing AI solutions for physics simulation and engineering can leverage smaller, faster models that generalize across different problem resolutions.

The research demonstrates that fundamental innovations in neural architecture design continue unlocking significant efficiency gains. Future work likely involves extending these techniques to higher-dimensional problems and evaluating performance on industrial-scale simulations. The discretization invariance property remains particularly valuable for practitioners needing models that adapt across different mesh resolutions without retraining.

• →SirenFNO eliminates frequency truncation in Fourier Neural Operators, enabling full-spectrum learning for PDEs with high-frequency components
• →Achieves 4-15x parameter reduction compared to baseline FNOs while maintaining discretization invariance across multiple benchmarks
• →Functional tensor decomposition variants reduce parameters by up to 73x, suggesting substantial practical deployment advantages
• →Addresses spectral bias limitations that hindered FNO performance on oscillatory PDEs critical to physics simulations
• →Constant parameter counts independent of grid resolution enable practical scaling for variable discretization scenarios