SSF-Net
A dual-branch neural network architecture for learning partial differential equation (PDE) operators. It fuses a parallel spatial branch (multi-directional Mamba) and a spectral branch (Fourier transform plus block-diagonal MLP) within SSF blocks, accommodating both the local propagation and global coupling dynamics of PDEs, and is embedded in a multi-resolution convolutional encoder-decoder.
Timeline (last 90 days)
The core of SSF-Net is a dual-branch spatial-spectral fusion block (SSF block), composed of a spatial branch and a spectral branch in parallel, where the spatial branch captures local interactions and the spectral branch extracts global features, and the two are fused within the block.
The SSF block is embedded in a multi-resolution convolutional encoder-decoder to extract multi-scale features.
The spatial branch of SSF-Net adopts a multi-directional Mamba structure to mitigate the directional bias introduced by the unidirectional causal scanning of standard Mamba, capturing local interactions while maintaining linear complexity.
The spectral branch of SSF-Net uses the Fourier transform to map features into the frequency domain, and then processes global spectral features through a block-diagonal MLP.
Compared with attention-based and convolution-based baseline models, SSF-Net achieves lower prediction error with fewer parameters.
SSF-Net maintains near-linear asymptotic complexity in the number of spatial tokens.
SSF-Net's differentiation lies in replacing the fully connected spectral processing commonly used in combinations of Mamba-type structures with Fourier methods by a block-diagonal MLP, thereby compressing the parameter count.
Public information about SSF-Net currently comes mainly from the paper itself, and independent third-party reproduction or comparative analysis has not yet appeared.
The specific performance of SSF-Net's long-horizon stability and its generalization ability on multi-physics coupled problems still require more experimental validation.
All Facts (9)
The core of SSF-Net is a dual-branch spatial-spectral fusion block (SSF block), composed of a spatial branch and a spectral branch in parallel, where the spatial branch captures local interactions and the spectral branch extracts global features, and the two are fused within the block.
50%UnverifiedThe SSF block is embedded in a multi-resolution convolutional encoder-decoder to extract multi-scale features.
50%UnverifiedThe spatial branch of SSF-Net adopts a multi-directional Mamba structure to mitigate the directional bias introduced by the unidirectional causal scanning of standard Mamba, capturing local interactions while maintaining linear complexity.
50%UnverifiedThe spectral branch of SSF-Net uses the Fourier transform to map features into the frequency domain, and then processes global spectral features through a block-diagonal MLP.
50%UnverifiedCompared with attention-based and convolution-based baseline models, SSF-Net achieves lower prediction error with fewer parameters.
50%UnverifiedSSF-Net maintains near-linear asymptotic complexity in the number of spatial tokens.
50%UnverifiedSSF-Net's differentiation lies in replacing the fully connected spectral processing commonly used in combinations of Mamba-type structures with Fourier methods by a block-diagonal MLP, thereby compressing the parameter count.
50%UnverifiedPublic information about SSF-Net currently comes mainly from the paper itself, and independent third-party reproduction or comparative analysis has not yet appeared.
50%UnverifiedThe specific performance of SSF-Net's long-horizon stability and its generalization ability on multi-physics coupled problems still require more experimental validation.
50%