Abstract:We propose an accurate numerical scheme for approximating the solution of the two dimensional acoustic wave problem. We use machine learning to find a stencil suitable even in the presence of high wavenumbers. The proposed scheme incorporates physically informed elements from the field of optimized numerical schemes into a convolutional optimization machine learning algorithm.
What problem does this paper attempt to address?
This paper attempts to solve the accuracy problem in the numerical solution of the two - dimensional acoustic wave equation in the case of low resolution (coarse grid) and high wavenumber. Specifically, the paper proposes a new numerical scheme, aiming to reduce the numerical dispersion error by optimizing the finite difference (FD) stencil, thereby improving the accuracy of the solution.
### Main problems:
1. **Numerical dispersion problem**: When using the finite difference method to solve the wave equation, numerical dispersion will cause the propagation speed of high - frequency components to be different from that of physical dispersion, resulting in errors. Especially in the case of high wavenumber and low resolution, this error is particularly significant.
2. **Limitations of traditional methods**:
- **Physical method**: Restricting the wavenumber range to reduce numerical dispersion, but valuable data will be lost.
- **Refining the grid**: Adding grid nodes can reduce numerical dispersion, but the computational cost is too high.
- **High - order accuracy method**: Using Taylor series to approximate low wavenumbers, but it has a poor effect on high wavenumbers.
- **Scheme for maintaining the dispersion relation (DRP)**: Existing DRP schemes perform well within a specific wavenumber range, but have limitations in broadband scenarios.
### Solutions:
The paper proposes a convolutional dispersion relation - preserving (CDPR) scheme based on convolution optimization and machine learning. The specific steps are as follows:
1. **Generate data set**: Create a data set of analytical solutions of the two - dimensional wave equation with different initial conditions, especially select those wavenumbers that show high numerical dispersion in classical numerical methods.
2. **Train the model**: Use a machine - learning model (neural network) to generate a 5×5 convolution kernel as an FD stencil and optimize it by minimizing the mean - squared - error (MSE) loss function.
3. **Physical information constraint**: Add a physical information (PI) loss function during the training process to ensure that the waveform predicted by the model is as close as possible to the analytical solution. In addition, the consistency, symmetry and fourth - order accuracy constraints of the stencil are also added to ensure the accuracy of the numerical solution.
4. **Test and verification**: Test the new scheme under multiple initial conditions and compare it with the classical second - order, fourth - order finite difference methods and existing DRP schemes to verify its superiority.
### Experimental results:
- **Low - wavenumber experiment**: The proposed fourth - order AI scheme (AI4) outperforms other methods under all discrete parameters; the second - order AI scheme (AI2) is also more accurate than the classical second - order FD and the benchmark DRP scheme.
- **High - wavenumber experiment**: AI4 shows the lowest error in the case of high wavenumber, especially its performance is significantly better than other methods under the finest grid. AI2 performs well under extremely coarse grids, but is not as good as AI4 under fine grids.
- **Ricker wave experiment**: Using the Ricker wave as an initial condition, the AI scheme performs well in numerical dispersion control, especially the second - order AI scheme has the best accuracy under coarser grids.
In conclusion, through the combination of machine learning and physical information optimization, this paper proposes an effective numerical scheme that can accurately solve the two - dimensional acoustic wave equation under low - resolution and high - wavenumber conditions and significantly reduce the numerical dispersion error.