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

Uladzislau Kapustsin,Utku Kaya,Thomas Richter

DOI: https://doi.org/10.1002/pamm.202300135

2023-07-03

Abstract:We analyze a hybrid method that enriches coarse grid finite element solutions with fine scale fluctuations obtained from a neural network. The idea stems from the Deep Neural Network Multigrid Solver (DNN-MG), (Margenberg et al., J Comput Phys 460:110983, 2022; A neural network multigrid solver for the Navier-Stokes equations) which embeds a neural network into a multigrid hierarchy by solving coarse grid levels directly and predicting the corrections on fine grid levels locally (e.g. on small patches that consist of several cells) by a neural network. Such local designs are quite appealing, as they allow a very good generalizability. In this work, we formalize the method and describe main components of the a-priori error analysis. Moreover, we numerically investigate how the size of training set affects the solution quality.

Numerical Analysis

What problem does this paper attempt to address?

### What problems does this paper attempt to solve?

This paper aims to solve the computational efficiency and accuracy problems encountered when using the combination of neural networks and the finite element method to solve partial differential equations (PDEs). Specifically, the paper proposes a hybrid finite element/neural network solver and applies it to the Poisson equation problem.

#### Main problems
1. **Efficiency of solving high - dimensional or parameter - dependent partial differential equations**:
- Although physics - inspired neural networks (PINNs) have advantages in terms of expressiveness and embedding of physical principles, they need to be retrained when parameters change, resulting in low efficiency.
- For classical problems (such as three - dimensional fluid dynamics), traditional discretization methods are far more efficient and accurate than neural network methods.

2. **Efficient solution of multi - scale problems**:
- In a multi - grid hierarchy, how to use neural networks to predict corrections on fine grids, thereby improving the accuracy of coarse - grid solutions.
- Specifically applied to the transient Navier - Stokes equation, a coarse solution is obtained through a classical finite element solver at each time step, and a fine - grid correction is locally predicted through a neural network.

3. **Error analysis and numerical experiments**:
- Formalize the hybrid method and describe the main components of its a priori error analysis.
- Numerically study the influence of the training set size on the solution quality.

#### Overview of solutions
The paper proposes a hybrid method that combines the coarse - grid finite element solution with the fine - scale fluctuations predicted by the neural network. This method is derived from the deep neural network multi - grid solver (DNN - MG), and its core idea is to embed a neural network in a multi - grid hierarchy, directly solve the coarse - grid layer and locally predict corrections at the fine - grid layer.

The specific steps are as follows:
- **Coarse - grid solution**: Use a classical finite element solver to obtain an initial solution on the coarse grid.
- **Fine - grid correction**: Locally predict fine - grid corrections through a neural network within a small area (i.e., a patch consisting of several elements).
- **Hybrid solution**: Combine the coarse - grid solution with the fine - grid corrections predicted by the neural network to form the final hybrid solution.

#### Application examples
The paper takes the two - dimensional Poisson equation as an example to show the application of this hybrid method. Numerical experiments are used to verify the influence of different training set sizes and grid refinement levels on the solution accuracy. The results show that, given sufficient data, this method can approach the fine solution used for training.

#### Formula summary
The weak form of the Poisson equation is:
\[
-\Delta u = f, \quad u|_{\partial \Omega} = 0,
\]
where \( f \in H^{-1}(\Omega) \) is the given force term.

The finite element space is defined as:
\[
V_h := \left\{ \phi \in C(\Omega) \mid \phi|_T \in P(r)(T), \forall T \in \mathcal{T}_h, \phi|_{\partial \Omega} = 0 \right\},
\]
where \( P(r)(T) \) is the \( r \) - degree polynomial space on the element \( T \).

The hybrid solution is defined as:
\[
u_N := u_H + \sum_{P \in \mathcal{T}_h} P_P w^P_N,
\]
where \( u_H \) is the coarse - grid solution and \( w^P_N \) is the correction predicted by the neural network on the local patch \( P \).

#### Conclusion
This paper solves the efficiency problems of traditional methods in dealing with high - dimensional or parameter - dependent partial differential equations by introducing a hybrid finite element/neural network solver, and verifies its effectiveness through numerical experiments.