Adaptive neural network basis methods for partial differential equations with low-regular solutions

Jianguo Huang,Haohao Wu,Tao Zhou
2024-11-04
Abstract:This paper aims to devise an adaptive neural network basis method for numerically solving a second-order semilinear partial differential equation (PDE) with low-regular solutions in two/three dimensions. The method is obtained by combining basis functions from a class of shallow neural networks and the resulting multi-scale analogues, a residual strategy in adaptive methods and the non-overlapping domain decomposition method. At the beginning, in view of the solution residual, we partition the total domain $\Omega$ into $K+1$ non-overlapping subdomains, denoted respectively as $\{\Omega_k\}_{k=0}^K$, where the exact solution is smooth on subdomain $\Omega_{0}$ and low-regular on subdomain $\Omega_{k}$ ($1\le k\le K$). Secondly, the low-regular solutions on different subdomains \(\Omega_{k}\)~($1\le k\le K$) are approximated by neural networks with different scales, while the smooth solution on subdomain \(\Omega_0\) is approximated by the initialized neural network. Thirdly, we determine the undetermined coefficients by solving the linear least squares problems directly or the nonlinear least squares problem via the Gauss-Newton method. The proposed method can be extended to multi-level case naturally. Finally, we use this adaptive method for several peak problems in two/three dimensions to show its high-efficient computational performance.
Numerical Analysis
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: propose an Adaptive Neural Network Basis method (ANNB) for the numerical solution of second - order semilinear partial differential equations (PDEs) with solutions of low regularity. Specifically, this method aims to deal with PDE problems with solutions of low regularity in two - or three - dimensional spaces. ### Core problems of the paper 1. **Challenges of solutions with low regularity**: - Traditional numerical methods often face difficulties when dealing with solutions of low regularity because these solutions may be very non - smooth in certain regions, leading to a decline in the accuracy of the numerical solution. 2. **Limitations of existing methods**: - Although deep learning methods (such as the deep Ritz method, weak adversarial networks, and physics - informed neural networks) perform well in solving high - dimensional PDEs, in low - dimensional cases, using shallow neural networks can better balance computational accuracy and cost. - When dealing with solutions of low regularity, existing methods may lead to a significant decline in accuracy due to the non - smoothness of the solutions. 3. **Objectives of the Adaptive Neural Network Basis method**: - Propose an adaptive method that combines shallow neural networks, non - overlapping domain decomposition methods, and multi - scale neural networks to improve the solution accuracy for solutions of low regularity. - Effectively deal with solutions of low regularity by dividing the total domain into multiple sub - domains and using neural networks of different scales on different sub - domains to approximate the solution. ### Main steps of the method 1. **Domain division**: - According to the residual of the solution, divide the total domain \(\Omega\) into \(K + 1\) non - overlapping sub - domains \(\{\Omega_k\}_{k = 0}^K\), where the solution on the sub - domain \(\Omega_0\) is smooth, while the solutions on the other sub - domains \(\Omega_k(1\leq k\leq K)\) are of low regularity. 2. **Neural network approximation**: - On different sub - domains \(\Omega_k(1\leq k\leq K)\), use neural networks of different scales to approximate the solutions of low regularity; while on the sub - domain \(\Omega_0\), use the initialized neural network to approximate the smooth solution. 3. **Determination of unknown coefficients**: - Determine the unknown coefficients in the neural network by solving the linear least - squares problem or the nonlinear least - squares problem (using the Gauss - Newton method). 4. **Multi - level extension**: - This method can be naturally extended to the multi - level case to further improve the solution accuracy. 5. **Numerical verification**: - Verify the high efficiency and high accuracy of this method through several two - dimensional and three - dimensional peak problems. ### Summary This paper proposes an innovative Adaptive Neural Network Basis method specifically for second - order partial differential equations with solutions of low regularity. By combining shallow neural networks, non - overlapping domain decomposition, and multi - scale neural networks, this method can significantly improve the solution accuracy for solutions of low regularity while ensuring computational efficiency.