Abstract:One of the oldest and most studied subject in scientific computing is algorithms for solving partial differential equations (PDEs). A long list of numerical methods have been proposed and successfully used for various applications. In recent years, deep learning methods have shown their superiority for high-dimensional PDEs where traditional methods fail. However, for low dimensional problems, it remains unclear whether these methods have a real advantage over traditional algorithms as a direct solver. In this work, we propose the random feature method (RFM) for solving PDEs, a natural bridge between traditional and machine learning-based algorithms. RFM is based on a combination of well-known ideas: 1. representation of the approximate solution using random feature functions; 2. collocation method to take care of the PDE; 3. the penalty method to treat the boundary conditions, which allows us to treat the boundary condition and the PDE in the same footing. We find it crucial to add several additional components including multi-scale representation and rescaling the weights in the loss function. We demonstrate that the method exhibits spectral accuracy and can compete with traditional solvers in terms of both accuracy and efficiency. In addition, we find that RFM is particularly suited for complex problems with complex geometry, where both traditional and machine learning-based algorithms encounter difficulties.

What problem does this paper attempt to address?

The main problems that this paper attempts to solve are the complementarity and limitations of traditional algorithms and machine - learning methods in solving partial differential equations (PDEs). Specifically: 1. **Complex geometry problems**: For problems with complex geometric shapes, although the traditional finite element method is theoretically very suitable for dealing with such problems, in practical applications, generating a suitable mesh is often a very time - consuming and complex task. While machine - learning - based methods are simple to code, their reliability and competitiveness have not yet been proven in practical applications. 2. **High - dimensional problems**: For high - dimensional PDEs, traditional algorithms usually cannot handle them effectively, while deep - learning methods have shown superiority in these high - dimensional problems. However, it is still unclear whether deep - learning methods are really superior to traditional algorithms for low - dimensional problems. 3. **Multiscale problems**: For example, problems involving chemical kinetics, which usually span multiple time scales; fully developed turbulent flows, which contain a large number of spatial and time scales; and the modeling of composite materials, etc. These problems pose challenges to both traditional algorithms and machine - learning methods. To solve the above problems, the author proposes the Random Feature Method (RFM), aiming to combine the advantages of traditional algorithms and machine - learning methods and propose a new method for solving PDEs. The key features of RFM include: - **Random feature functions**: Use random feature functions to represent approximate solutions, which makes the method more flexible in dealing with complex geometry problems. - **Collocation point method**: Use the collocation point method to deal with PDEs and their boundary conditions in the form of the least - squares method. - **Penalty method**: Use the penalty method to deal with boundary conditions, making them treated at the same level as PDEs. - **Multiscale representation**: Introduce multiscale representation, use the Partition of Unity (PoU) method to stitch different local representations together and provide a global representation. - **Weight rescaling**: Adjust the weight parameters of each part in the loss function to balance the contributions of PDEs and boundary conditions. Through these methods, RFM can achieve spectral accuracy and perform excellently in dealing with complex geometry problems, and can compete with traditional solvers in terms of accuracy and efficiency.

Bridging Traditional and Machine Learning-based Algorithms for Solving PDEs: The Random Feature Method

Optimization of Random Feature Method in the High-Precision Regime

The Random Feature Method for Time-dependent Problems

The Random Feature Method for Solving Interface Problems

An Extreme Learning Machine-Based Method for Computational PDEs in Higher Dimensions

Solving PDEs on unknown manifolds with machine learning

High-precision randomized iterative methods for the random feature method

Algorithms for Solving High Dimensional PDEs: from Nonlinear Monte Carlo to Machine Learning

Non-equispaced Fourier Neural Solvers for PDEs

Monte Carlo Neural PDE Solver for Learning PDEs via Probabilistic Representation

Solving multiscale elliptic problems by sparse radial basis function neural networks

Convergence Analysis of a Quasi-Monte Carlo-based Deep Learning Algorithm for Solving Partial Differential Equations

A novel paradigm for solving PDEs: multi scale neural computing

A deep learning solution approach for high-dimensional random differential equations

Solving Partial Differential Equations Using Deep Learning and Physical Constraints

Deep-learning of parametric partial differential equations from sparse and noisy data

MRF-PINN: A Multi-Receptive-Field convolutional physics-informed neural network for solving partial differential equations

RBF-MGN:Solving spatiotemporal PDEs with Physics-informed Graph Neural Network

Low-rank approximation for multiscale PDEs

Local Extreme Learning Machines and Domain Decomposition for Solving Linear and Nonlinear Partial Differential Equations