Abstract
Lagrangian fluid simulations, which represent fluids as interacting particles, present significant challenges for machine learning due to the need for physical consistency and stability. We present Symmetric Fourier Basis Convolutions (SFBC), a framework that leverages separable Fourier basis functions to learn continuous convolutions with "baked-in" physical symmetries. By explicitly modeling even and odd interactions, SFBC achieves superior accuracy and stability compared to standard graph neural networks and linear convolutions, while eliminating the need for manual window functions. Our results show a 30% reduction in divergence error and significant parameter efficiency, requiring an order of magnitude fewer parameters than state-of-the-art graph simulators.
Introduction
Fluid dynamics is a notoriously difficult field for machine learning. Whether modeling the swirling vortices of a smoke plume or the high-pressure impact of a breaking wave, physical simulations require a delicate balance of stability, accuracy, and conservation laws. In Lagrangian mechanics, fluids are represented as a collection of moving particles. To predict the evolution of these systems, researchers often employ Continuous Convolutions (CConvs), which calculate interactions based on the relative positions of neighboring particles.
However, standard CConvs suffer from two primary weaknesses. First, they lack strong inductive biases; most models use simple linear interpolation or Multi-Layer Perceptrons (MLPs) to learn interaction "shapes," which often fail to capture the inherent smoothness and symmetry required by physics. Second, they rely on "window functions"—mathematical tapers that force interactions to zero at a certain distance—which are typically hand-chosen and sub-optimal for learning.
Methodology: Symmetric Fourier Basis Convolutions
Presented at ICLR 2024 by Rene Winchenbach and Nils Thuerey, Symmetric Fourier Basis Convolutions (SFBC) address these issues by using a Fourier series as a basis for the convolutional kernel.
Separable Basis Functions
The core innovation lies in constructing 2D or 3D kernels as an outer product of 1D basis functions. This separability makes the computation highly efficient and allows the network to learn complex, smooth shapes with very few parameters. By using a Fourier basis, the model naturally learns smooth, periodic-like interactions that represent the underlying physics of Smoothed Particle Hydrodynamics (SPH) more effectively than piece-wise linear functions.
Symmetry by Design
A crucial aspect of SFBC is the explicit inclusion of even and odd symmetry terms. Different physical quantities behave differently: for instance, density is typically a symmetric (even) interaction, while pressure gradients are anti-symmetric (odd). By baking these symmetries into the architecture, SFBC ensures that the learned "gradient" of a simulation is fundamentally more accurate and stable than models that are "physics-blind."
Experimental Evaluation
The authors evaluated SFBC on four distinct scenarios designed to test various physical properties:
- Toy Problems: Learning simple SPH kernels and their gradients to verify fundamental capabilities.
- Compressible 1D SPH: Testing the model's ability to handle varied particle distributions.
- Weakly-Compressible 2D SPH: Focusing on long-term inference stability and the preservation of vortices in a closed domain.
- Incompressible 2D SPH: A "fluid blob collision" scenario featuring free surfaces, serving as a classic stress test for particle solvers.
Key Results and Discussion
SFBC demonstrates a remarkable combination of accuracy, generalization, and efficiency.
Accuracy and Stability
In 2D stability tests, SFBC reduced the divergence error—a key measure of how well a fluid's incompressibility is maintained—by 30% compared to the previous state-of-the-art method, LinCConv. The paper demonstrates that models failing to account for even/odd symmetries are fundamentally less stable.
Parameter Efficiency
While MLPs are universal approximators, they lack the physical priors inherent in Fourier bases. SFBC achieved superior results with only 30,000 parameters, whereas general Graph Network Simulators (GNS) required nearly 400,000 parameters to reach comparable performance. Despite the mathematical complexity, the computational overhead is minimal, with only a 0.5% to 12% increase in runtime over simpler linear convolutions.
The End of Window Functions
Perhaps the most surprising finding is that when using a Fourier basis, traditional SPH window functions (like the Müller or Spiky kernels) become redundant. The Fourier basis is expressive enough to learn its own optimal "tapering" at the edges of the interaction radius, simplifying the simulation pipeline.
Conclusion
Symmetric Fourier Basis Convolutions represent a significant advancement in learning Lagrangian fluid mechanics. By prioritizing physical symmetry and smoothness, the framework provides a robust, efficient, and highly accurate alternative to traditional neural simulation methods. While future work is needed to explore parameter scaling and initialization for other basis types like Chebyshev polynomials, the framework's generality suggests potential applications in molecular dynamics and astrophysics.
Summary updated