Tamara G. Grossmann,Urszula Julia Komorowska,Jonas Latz,Carola-Bibiane Schönlieb
Abstract:Partial differential equations play a fundamental role in the mathematical modelling of many processes and systems in physical, biological and other sciences. To simulate such processes and systems, the solutions of PDEs often need to be approximated numerically. The finite element method, for instance, is a usual standard methodology to do so. The recent success of deep neural networks at various approximation tasks has motivated their use in the numerical solution of PDEs. These so-called physics-informed neural networks and their variants have shown to be able to successfully approximate a large range of partial differential equations. So far, physics-informed neural networks and the finite element method have mainly been studied in isolation of each other. In this work, we compare the methodologies in a systematic computational study. Indeed, we employ both methods to numerically solve various linear and nonlinear partial differential equations: Poisson in 1D, 2D, and 3D, Allen-Cahn in 1D, semilinear Schr\"odinger in 1D and 2D. We then compare computational costs and approximation accuracies. In terms of solution time and accuracy, physics-informed neural networks have not been able to outperform the finite element method in our study. In some experiments, they were faster at evaluating the solved PDE.
What problem does this paper attempt to address?
### Problems Addressed by the Paper
The main purpose of this paper is to systematically compare the performance of Physics-Informed Neural Networks (PINNs) and the Finite Element Method (FEM) in solving Partial Differential Equations (PDEs). Specifically, the paper compares the two methods through numerical experiments on the following types of PDEs:
1. **Poisson equation**:
- 1D, 2D, and 3D elliptic Poisson equations.
2. **Allen-Cahn equation**:
- 1D parabolic Allen-Cahn equation.
3. **Semilinear Schrödinger equation**:
- 1D and 2D hyperbolic semilinear Schrödinger equations.
The paper compares the two methods in terms of solving time, evaluation time, and accuracy. To conduct a systematic comparison, the researchers first determined the true solutions for each type of equation. For the Poisson equation, analytical solutions were used as the true values; for the Allen-Cahn equation and the semilinear Schrödinger equation, very fine finite element solutions were used as reference true values.
### Method Overview
#### Experimental Design
- **Finite Element Method (FEM)**: Different mesh sizes were used to solve the equations, and the computation time and accuracy for different mesh sizes were analyzed. For time-dependent equations (such as the Allen-Cahn equation and the semilinear Schrödinger equation), a semi-implicit strategy was used for discretization.
- **Physics-Informed Neural Networks (PINNs)**: Following the original PINNs method, a loss function was designed, and the Adam optimizer was used for coarse optimization, followed by the L-BFGS optimizer for fine-tuning. During training, Latin hypercube sampling points were resampled once per epoch to cover more sampling space.
#### Comparison Metrics
- **Solving Time**: The time required for each method to solve a general solution.
- **Evaluation Time**: For FEM, the time required to interpolate the solution to different meshes; for PINNs, the time required to evaluate the trained neural network at new data points.
- **Accuracy**: The accuracy of both methods relative to the true solution measured on the same mesh, using the ℓ2 relative error as the accuracy metric.
### Conclusion
By systematically comparing the performance of the Finite Element Method and Physics-Informed Neural Networks on different types of PDEs, the paper finds that while PINNs are faster in evaluation in some experiments, they do not surpass FEM in overall solving time and accuracy. These results provide valuable insights into the strengths and weaknesses of both methods in practical applications.