Abstract:In this paper, a new weighted first-order formulation is proposed for solving the anisotropic diffusion equations with deep neural networks. For many numerical schemes, the accurate approximation of anisotropic heat flux is crucial for the overall accuracy. In this work, the heat flux is firstly decomposed into two components along the two eigenvectors of the diffusion tensor, thus the anisotropic heat flux approximation is converted into the approximation of two isotropic components. Moreover, to handle the possible jump of the diffusion tensor across the interface, the weighted first-order formulation is obtained by multiplying this first-order formulation by a weighted function. By the decaying property of the weighted function, the weighted first-order formulation is always well-defined in the pointwise way. Finally, the weighted first-order formulation is solved with deep neural network approximation. Compared to the neural network approximation with the original second-order elliptic formulation, the proposed method can significantly improve the accuracy, especially for the discontinuous anisotropic diffusion problems.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: when solving the diffusion equation with discontinuous anisotropic diffusion tensors, traditional numerical methods (such as the finite - volume method) encounter the problem of decreasing accuracy when dealing with complex geometries and discontinuous interfaces. Specifically, when there are material discontinuities or highly discontinuous diffusion tensors caused by local physical fields (such as magnetic fields), traditional discretization methods are difficult to accurately capture the direction and magnitude of heat flux, especially on discontinuous interfaces. To solve this problem, the author proposes a new weighted first - order formulation based on deep neural networks. By decomposing the heat flux into two components along the eigenvectors of the diffusion tensor and introducing a decay weight function to handle discontinuous interfaces, this method can significantly improve the solution accuracy for discontinuous anisotropic diffusion problems. Compared with the neural network approximation method that directly uses the original second - order elliptic formulation, the new method shows higher accuracy when dealing with discontinuity problems. ### Main improvement points: 1. **Heat flux decomposition**: Decompose the heat flux into two components along the eigenvectors of the diffusion tensor, thereby converting the approximation of anisotropic heat flux into the approximation of two isotropic components. 2. **Weighted first - order formula**: By introducing a decay weight function, the first - order formula near the discontinuous interface can still be well - defined, avoiding the accuracy loss of traditional methods at discontinuities. 3. **Deep neural network solution**: Use deep neural networks to solve the weighted first - order formula, improving the ability to solve problems with complex geometries and discontinuous interfaces. ### Advantages: - For both smooth and non - smooth anisotropic diffusion problems, this method can maintain high solution accuracy. - When dealing with discontinuous interfaces, it can better capture the changes in heat flux and avoid the error amplification of traditional methods at discontinuities. - The method is applicable to two - dimensional and three - dimensional problems, and performs well in high - dimensional cases without being affected by the curse of dimensionality. In conclusion, the method proposed in this paper aims to provide a more accurate and flexible solution to anisotropic diffusion problems by combining deep learning and the weighted first - order formula, especially in the presence of discontinuous interfaces.

A weighted first-order formulation for solving anisotropic diffusion equations with deep neural networks

Analysis and Design of Anisotropic Diffusion for Image Processing

Simulating anisotropic diffusion processes with smoothed particle hydrodynamics

Solving Euler equations with gradient-weighted multi-input high-dimensional feature neural network

An Asymptotic Preserving method for strongly anisotropic diffusion equations based on field line integration

A deep neural network approach on solving the linear transport model under diffusive scaling

Optimally weighted loss functions for solving PDEs with Neural Networks

A physics-inspired neural network to solve partial differential equations - application in diffusion-induced stress

Physics-Informed Neural Networks for 1-D Steady-State Diffusion-Advection-Reaction Equations

Cell-average based neural network method for hyperbolic and parabolic partial differential equations

A deep learning method for multi-material diffusion problems based on physics-informed neural networks

An improved physics-informed neural network with adaptive weighting and mixed differentiation for solving the incompressible Navier–Stokes equations

Deep neural network methods for solving forward and inverse problems of time fractional diffusion equations with conformable derivative

Deep Neural Network Approach to Forward-Inverse Problems

A physics-informed neural network based on mixed data sampling for solving modified diffusion equations

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

Energetic Variational Neural Network Discretizations of Gradient Flows

A discretization-free deep neural network-based approach for advection-dispersion-reaction mechanisms

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

A feedforward neural network framework for approximating the solutions to nonlinear ordinary differential equations

Higher-Order Generalized Finite Differences for Variable Coefficient Diffusion Operators