Abstract
Partial differential equations (PDEs) are widely used across the physical and computational sciences. Decades of research and engineering went into designing fast iterative solution methods. Existing solvers are general purpose, but may be sub-optimal for specific classes of problems. In contrast to existing hand-crafted solutions, we propose an approach to learn a fast iterative solver tailored to a specific domain. We achieve this goal by learning to modify the updates of an existing solver using a deep neural network. Crucially, our approach is proven to preserve strong correctness and convergence guarantees. After training on a single geometry, our model generalizes to a wide variety of geometries and boundary conditions, and achieves 2-3 times speedup compared to state-of-the-art solvers.
Introduction
Partial Differential Equations (PDEs) are the mathematical backbone of physics, describing phenomena from heat flow to fluid dynamics. Traditionally, these are solved using hand-crafted iterative methods like Jacobi or Multigrid. While reliable, these general-purpose solvers are often sub-optimal for specific problem classes. This paper introduces a method to "supercharge" traditional solvers by using deep neural networks to learn domain-specific update rules, providing rigorous mathematical guarantees for convergence and correctness while achieving significant speedups.
Methodology: Learning Neural Updates
Instead of replacing traditional solvers with a "black-box" neural network, the authors propose a hybrid approach. They modify a standard iterative solver update $\Psi(u)$ into a new learned iterator, $\Phi_H$: $$\Phi_H(u) = \Psi(u) + GHw$$ Where:
- $\Psi(u)$ is the standard, "safe" traditional update.
- $w$ is the difference between the new and old solution.
- $H$ is a learned linear operator parameterized as a Deep Neural Network (using Conv or U-Net architectures).
- $G$ is a "reset" matrix that ensures boundary conditions are strictly maintained.
This specific structure ensures that the neural solver always converges to the same correct "fixed point" as the original solver, even if the neural network is poorly trained.
Experimental Setup and Generalization
The researchers focused on the 2D Poisson equation. The model was trained on a simple $64 \times 64$ square domain with the Laplace equation ($\nabla^2 u = 0$). To test generalization, they evaluated the model on much more complex scenarios, including larger grids, different geometries, and non-zero source terms.
By using convolutional layers that focus on local neighborhoods, the model learns the local physics of the PDE. This allows it to work on complex, curved geometries it has never seen during training.
Key Results
- Speedup: The neural solvers achieved a 2-3x speedup in iteration count compared to standard Jacobi and Multigrid baselines.
- Hardware Efficiency: Built with standard neural network layers, the solver is natively optimized for GPUs, leading to a 30x speedup over CPU-based traditional solvers.
- Zero-Shot Generalization: The model maintained its performance gains across different geometries (L-shapes, cylinders) and non-zero source terms without any additional training.
Conclusion
This framework demonstrates that deep learning can be integrated into traditional numerical methods to provide significant performance gains without sacrificing mathematical rigor. While currently limited to linear PDEs on fixed meshes, the success of this "learned correction" approach opens new avenues for accelerating complex physical simulations in engineering and science.
Summary updated