Abstract:We show how to construct the deep neural network (DNN) expert to predict quasi-optimal $hp$-refinements for a given computational problem. The main idea is to train the DNN expert during executing the self-adaptive $hp$-finite element method ($hp$-FEM) algorithm and use it later to predict further $hp$ refinements. For the training, we use a two-grid paradigm self-adaptive $hp$-FEM algorithm. It employs the fine mesh to provide the optimal $hp$ refinements for coarse mesh elements. We aim to construct the DNN expert to identify quasi-optimal $hp$ refinements of the coarse mesh elements. During the training phase, we use the direct solver to obtain the solution for the fine mesh to guide the optimal refinements over the coarse mesh element. After training, we turn off the self-adaptive $hp$-FEM algorithm and continue with quasi-optimal refinements as proposed by the DNN expert trained. We test our method on three-dimensional Fichera and two-dimensional L-shaped domain problems. We verify the convergence of the numerical accuracy with respect to the mesh size. We show that the exponential convergence delivered by the self-adaptive $hp$-FEM can be preserved if we continue refinements with a properly trained DNN expert. Thus, in this paper, we show that from the self-adaptive $hp$-FEM it is possible to train the DNN expert the location of the singularities, and continue with the selection of the quasi-optimal $hp$ refinements, preserving the exponential convergence of the method.

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is to predict the quasi - optimal hp - finite - element refinement strategy through deep neural networks (DNN) in order to reduce the consumption of computational resources and maintain the exponential convergence of numerical solutions. Specifically, the researchers hope to develop a DNN expert system that can replace the traditional two - grid paradigm of coarse and fine grids in the adaptive hp - finite - element method (hp - FEM) to select quasi - optimal refinement schemes. ### Main Problem Description 1. **Limitations of Traditional Adaptive hp - FEM**: - The adaptive hp - FEM algorithm relies on solving problems on coarse and fine grids and comparing the errors between them to determine the optimal refinement strategy. - Although this method can provide an exponential convergence rate, the computational cost is very high, especially in three - dimensional problems, where the size of the fine - grid problem is usually two orders of magnitude larger than that of the coarse - grid. - In addition, due to the high non - uniformity of the grid and the difference in polynomial orders of different elements, the iterative solver is difficult to converge in these cases, so a computationally expensive direct solver is usually required. 2. **Purpose of Introducing DNN**: - Train a DNN expert system, collect data during the execution of the adaptive hp - FEM algorithm, and use this data to train the DNN. - Once the DNN is trained, the adaptive hp - FEM algorithm can be turned off and the DNN expert system can be used instead to predict subsequent quasi - optimal refinement schemes. - The inputs of the DNN include the position, size of elements and the current polynomial order, and the output is a decision on how to perform refinement (h - refinement or p - refinement). - In this way, the DNN can learn the singularity positions and how to perform appropriate refinement at these positions, so that it can continue to refine without invoking the solver. ### Key Points of the Solution - **Data Collection and Training**: During the operation of the adaptive hp - FEM algorithm, collect all data on optimal refinement decisions as the training set for the DNN. - **DNN Architecture Design**: Design a feed - forward neural network (feed - forward DNN) with multiple branches to handle h - refinement and p - refinement decisions respectively. - **Application and Verification**: Apply the trained DNN to new problem instances, verify whether it can maintain exponential convergence, and compare it with the traditional adaptive hp - FEM algorithm. ### Experimental Verification The paper verifies the effectiveness of the proposed method through two model problems: 1. **Two - Dimensional L - Shaped Domain Problem**: - It is verified that the performance of the DNN - driven hp - FEM algorithm in the last 10 iterations is similar to that of the adaptive hp - FEM algorithm, and both can maintain exponential convergence. 2. **Three - Dimensional Fichera Problem**: - Train the DNN with the data generated by the 12 - iteration adaptive hp - FEM algorithm, and verify whether the DNN can correctly predict the singularity positions and refinement patterns. In general, this paper shows how to use deep neural networks to assist the adaptive hp - finite - element method, thereby maintaining efficient numerical solution convergence performance while reducing the consumption of computational resources.

Quasi-optimal $hp$-finite element refinements towards singularities via deep neural network prediction

Adaptive Finite Element Interpolated Neural Networks

Adaptive Finite Element Method for a Nonlinear Helmholtz Equation with High Wave Number

HiDeNN-FEM: a seamless machine learning approach to nonlinear finite element analysis

A Stable Generalized Finite Element Method Coupled with Deep Neural Network for Interface Problems with Discontinuities

Hierarchical Deep-Learning Neural Networks: Finite Elements and Beyond

Finite Element Method-enhanced Neural Network for Forward and Inverse Problems

HNS: An efficient hermite neural solver for solving time-fractional partial differential equations

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

Neural network-augmented differentiable finite element method for boundary value problems

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

Convergence and Optimality of hp-AFEM

Self-Adaptive Strategy for One-Dimensional Finite Element Method Based on Eep Method with Optimal Super-Convergence Order

Error analysis for hybrid finite element/neural network discretizations

Hybrid FEM-NN models: Combining artificial neural networks with the finite element method

Finite Element Neural Network Interpolation. Part II: Hybridisation with the Proper Generalised Decomposition for non-linear surrogate modelling

Compatible finite element interpolated neural networks

Finite Element Neural Network Interpolation. Part I: Interpretable and Adaptive Discretization for Solving PDEs

Physics-informed neural networks with parameter asymptotic strategy for learning singularly perturbed convection-dominated problem

On full linear convergence and optimal complexity of adaptive FEM with inexact solver

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