Abstract:In this paper, we introduce cell-average based neural network (CANN) method to solve high-dimensional parabolic partial differential equations. The method is based on the integral or weak formulation of partial differential equations. A feedforward network is considered to train the solution average of cells in neighboring time. Initial values and approximate solution at $t=\Delta t$ obtained by high order numerical method are taken as the inputs and outputs of network, respectively. We use supervised training combined with a simple backpropagation algorithm to train the network parameters. We find that the neural network has been trained to optimality for high-dimensional problems, the CFL condition is not strictly limited for CANN method and the trained network is used to solve the same problem with different initial values. For the high-dimensional parabolic equations, the convergence is observed and the errors are shown related to spatial mesh size but independent of time step size.
What problem does this paper attempt to address?
This paper attempts to solve the numerical solution problem of high - dimensional parabolic partial differential equations (PDEs). Specifically, the author proposes a cell - average - based neural network method (CANN) to overcome the computational complexity and time - step limitations encountered by traditional numerical methods when dealing with high - dimensional problems.
### Core problems of the paper
1. **Challenges of high - dimensional problems**:
- High - dimensional partial differential equations are widely used in fields such as physics, engineering, and finance.
- Traditional numerical methods (such as the finite - difference method, the finite - volume method, etc.) face the "curse of dimensionality" problem in high - dimensional situations, that is, as the dimension increases, the number of grid points grows exponentially, resulting in a sharp increase in computational cost.
- Explicit difference schemes are strictly limited by the CFL condition, and the time step must be very small, further increasing the computational burden.
2. **Advantages of neural networks**:
- Neural network models perform well in solving high - dimensional deterministic differential equations. They can perform differentiation automatically, do not require meshing, and can be applied to various types of equations.
- The CANN method combines the ideas of neural networks and the finite - volume method. By training the neural network to approximate the change in cell - average values, it solves the problem of solving high - dimensional parabolic equations.
### Main contributions
- Propose a new cell - average - based neural network method (CANN) for solving high - dimensional parabolic partial differential equations.
- This method uses the integral or weak form of partial differential equations and combines supervised learning and a simple back - propagation algorithm to train neural network parameters.
- Experimental results show that the CANN method can not only effectively solve high - dimensional problems but also has no strict limitation on the time step and is applicable to the same problem under different initial conditions.
### Formula representation
The main formulas involved in the paper include:
- General form of high - dimensional parabolic partial differential equations:
\[
u_t+\beta\nabla\cdot f(u)=\mu\Delta u + g(u),\quad(x,t)\in\Omega\times\mathbb{R}^+
\]
where \(t\) and \(x\) represent time and space variables respectively, \(\Omega\in\mathbb{R}^d\) is the spatial domain, \(d\geq2\).
- Definition of cell - average values:
\[
\bar{u}_{ij}(t)=\frac{1}{\Delta S}\int_{\Gamma_{ij}}u(x,y,t)\,dx\,dy
\]
- Squared loss function of neural network output and target output:
\[
L^{t_{n + 1}}(\Theta)=\sum_{i = 1}^I\sum_{j = 1}^J(\bar{v}_{ij}^{out}-\bar{u}_{ij}^{n+1})^2\Delta S
\]
Through these methods and techniques, the paper successfully solves the numerical solution problem of high - dimensional parabolic partial differential equations, providing new ideas and tools for research in related fields.