Abstract:Neural networks suffer from spectral bias having difficulty in representing the high frequency components of a function while relaxation methods can resolve high frequencies efficiently but stall at moderate to low frequencies. We exploit the weaknesses of the two approaches by combining them synergistically to develop a fast numerical solver of partial differential equations (PDEs) at scale. Specifically, we propose HINTS, a hybrid, iterative, numerical, and transferable solver by integrating a Deep Operator Network (DeepONet) with standard relaxation methods, leading to parallel efficiency and algorithmic scalability for a wide class of PDEs, not tractable with existing monolithic solvers. HINTS balances the convergence behavior across the spectrum of eigenmodes by utilizing the spectral bias of DeepONet, resulting in a uniform convergence rate and hence exceptional performance of the hybrid solver overall. Moreover, HINTS applies to large-scale, multidimensional systems, it is flexible with regards to discretizations, computational domain, and boundary conditions.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the efficiency and applicability issues encountered by existing partial differential equation (PDE) numerical solvers when dealing with large - scale, multi - dimensional systems. Specifically: 1. **Spectral Bias of Neural Networks**: Neural networks have difficulty representing the high - frequency components of functions, resulting in poor performance when dealing with high - frequency patterns. 2. **Limitations of Relaxation Methods**: Traditional relaxation methods (such as the Jacobi method and the Gauss - Seidel method) can efficiently analyze high - frequency components, but converge slowly or even stagnate on medium - and low - frequency components. 3. **Challenges in Solving Large - Scale Linear Systems**: Existing numerical solvers have low computational efficiency when dealing with large - scale linear systems and have convergence problems for asymmetric and indefinite systems. To solve these problems, the author proposes a new Hybrid Iterative Numerical Transport Solver (HINTS), which combines the Deep Operator Network (DeepONet) and standard relaxation methods, aiming to achieve a fast, accurate, and widely applicable PDE solver. The main features of HINTS include: - **Spectral Equalization**: Utilize the spectral bias characteristics of DeepONet to balance the convergence behavior of different frequency patterns, thereby achieving a uniform convergence rate. - **Flexibility**: Applicable to large - scale, multi - dimensional systems, and highly flexible with respect to discretization methods, computational domains, and boundary conditions. - **Pre - processing Ability**: Can be used as a pre - processor for Krylov methods to improve their convergence and algorithm scalability. Through a series of numerical experiments, the paper demonstrates the superior performance of HINTS in solving typical PDE problems such as the Poisson equation and the Helmholtz equation, especially showing significant advantages when dealing with indefinite systems and large - scale problems.

Blending Neural Operators and Relaxation Methods in PDE Numerical Solvers

Blending neural operators and relaxation methods in PDE numerical solvers

Non-equispaced Fourier Neural Solvers for PDEs

Quantifying Training Difficulty and Accelerating Convergence in Neural Network-Based PDE Solvers

Finite Operator Learning: Bridging Neural Operators and Numerical Methods for Efficient Parametric Solution and Optimization of PDEs

A Stable and Scalable Method for Solving Initial Value PDEs with Neural Networks

dNNsolve: an efficient NN-based PDE solver

Interfacing finite elements with deep neural operators for fast multiscale modeling of mechanics problems

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

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

Solving forward and inverse PDE problems on unknown manifolds via physics-informed neural operators

Deep Equilibrium Based Neural Operators for Steady-State PDEs

A novel paradigm for solving PDEs: multi scale neural computing

Operator Learning Enhanced Physics-informed Neural Networks for Solving Partial Differential Equations Characterized by Sharp Solutions

Neural Q-learning for solving PDEs

Message Passing Neural PDE Solvers

Synergistic Learning with Multi-Task DeepONet for Efficient PDE Problem Solving

Large scale scattering using fast solvers based on neural operators

A hybrid physics-informed neural network based multiscale solver as a partial differential equation constrained optimization problem

Enhancing Solutions for Complex PDEs: Introducing Complementary Convolution and Equivariant Attention in Fourier Neural Operators

Solving a class of multi-scale elliptic PDEs by means of Fourier-based mixed physics informed neural networks