Abstract:We introduce a neural-preconditioned iterative solver for Poisson equations with mixed boundary conditions. Typical Poisson discretizations yield large, ill-conditioned linear systems. Iterative solvers can be effective for these problems, but only when equipped with powerful preconditioners. Unfortunately, effective preconditioners like multigrid require costly setup phases that must be re-executed every time domain shapes or boundary conditions change, forming a severe bottleneck for problems with evolving boundaries. In contrast, we present a neural preconditioner trained to efficiently approximate the inverse of the discrete Laplacian in the presence of such changes. Our approach generalizes to domain shapes, boundary conditions, and grid sizes outside the training set. The key to our preconditioner's success is a novel, lightweight neural network architecture featuring spatially varying convolution kernels and supporting fast inference. We demonstrate that our solver outperforms state-of-the-art methods like algebraic multigrid as well as recently proposed neural preconditioners on challenging test cases arising from incompressible fluid simulations.

What problem does this paper attempt to address?

This paper attempts to solve the computational challenges encountered when solving the Poisson equation under mixed Dirichlet and Neumann boundary conditions. Specifically, the author aims to overcome the following problems: 1. **Ill - conditioning of large - scale linear systems**: The discretized Poisson equation usually results in large, ill - conditioned linear systems. These problems require efficient pre - processing methods to accelerate the solution. 2. **Limitations of traditional pre - processing methods**: Traditional pre - processing methods (such as the multigrid method) are effective, but when the domain shape or boundary conditions change, a costly setup phase needs to be re - executed, which forms a serious bottleneck in dealing with problems of evolving boundaries. 3. **Challenges in supporting mixed boundary conditions**: Many existing methods can only handle specific types of boundary conditions (such as pure Neumann boundary conditions) and cannot handle cases where Dirichlet and Neumann boundary conditions are present simultaneously very well. To solve these problems, the author proposes a neural - preconditioned solver, a pre - processing method based on neural networks, which can efficiently approximate the inverse of the discrete Laplacian matrix and can adapt to changing domain shapes and boundary conditions. Specific contributions include: - **Light - weight neural network architecture**: A novel architecture using spatially - varying convolution kernels can effectively approximate the inverse of the Laplacian matrix of a structured grid with arbitrary mixed Dirichlet and Neumann boundary conditions. - **Unsupervised training**: Unsupervised training is carried out through a simple residual loss function, allowing the network to generalize to systems not seen during the training process. - **Performance improvement**: In a series of challenging fluid simulation test cases, compared with state - of - the - art solvers (such as algebraic multigrid and incomplete Cholesky decomposition) and recently proposed neural pre - processing methods (such as DCDM and FluidNet), this method significantly improves performance. ### Main innovation points - **Support for mixed boundary conditions**: The new method is not only applicable to pure Neumann boundary conditions, but also can handle more complex mixed Dirichlet and Neumann boundary conditions, which is particularly important for problems such as simulating free - surface liquid flow. - **Fast information propagation**: Through a multi - resolution network architecture, information can be propagated faster on the computational grid, thereby improving the solution efficiency. - **Utilization of geometric information**: Custom convolution blocks (CConv) are introduced, which adjust the convolution weights according to local geometric information to better simulate the smoothing operation in the multigrid method. In conclusion, this paper proposes a novel neural pre - processing method, which solves the bottlenecks encountered by traditional methods when dealing with the Poisson equation under mixed boundary conditions, especially in the case of evolving boundary problems.

A Neural-preconditioned Poisson Solver for Mixed Dirichlet and Neumann Boundary Conditions

A Poisson’s Equation Solver Based on Neural Network Precondtioned CG Method

A hybrid finite element/neural network solver and its application to the Poisson problem

Non-equispaced Fourier Neural Solvers for PDEs

Learning Preconditioner for Conjugate Gradient PDE Solvers.

Solving the Poisson equation on small aspect ratio domains using unstructured meshes

A Spatial-Temporally Adaptive PINN Framework for 3D Bi-Ventricular Electrophysiological Simulations and Parameter Inference

Deep Learning-Enhanced Preconditioning for Efficient Conjugate Gradient Solvers in Large-Scale PDE Systems

A Neural Network-Based Poisson Solver for Fluid Simulation

Neural-Network-Assisted Finite Difference Discretization for Numerical Solution of Partial Differential Equations

An efficient neural-network and finite-difference hybrid method for elliptic interface problems with applications

A fully differentiable GNN-based PDE Solver: With Applications to Poisson and Navier-Stokes Equations

A hybrid iterative method based on MIONet for PDEs: Theory and numerical examples

An Adaptive Preconditioner for Three-Dimensional Single-Phase Compressible Flow in Highly Heterogeneous Porous Media

Semi-Implicit Neural Solver for Time-dependent Partial Differential Equations

PINN-MG: A Multigrid-Inspired Hybrid Framework Combining Iterative Method and Physics-Informed Neural Networks

Monte Carlo Neural PDE Solver for Learning PDEs via Probabilistic Representation

Multigrid-Augmented Deep Learning Preconditioners for the Helmholtz Equation using Compact Implicit Layers

Neural Poisson Solver: A Universal and Continuous Framework for Natural Signal Blending

Neural Green's Function Accelerated Iterative Methods for Solving Indefinite Boundary Value Problems

A mixed formulation for physics-informed neural networks as a potential solver for engineering problems in heterogeneous domains: Comparison with finite element method