Abstract:We present partial evolutionary tensor neural networks (pETNNs), a novel framework for solving time-dependent partial differential equations with high accuracy and capable of handling high-dimensional problems. Our architecture incorporates tensor neural networks and evolutional parametric approximation. A posterior error bounded is proposed to support the extrapolation capabilities. In the numerical implementations, we adopt a partial update strategy to achieve a significant reduction in computational cost while maintaining precision and robustness. Notably, as a low-rank approximation method of complex dynamical systems, the pETNNs enhance the accuracy of evolutional deep neural networks and empowers computational abilities to address high-dimensional problems. Numerical experiments demonstrate the superior performance of the pETNNs in solving time-dependent complex equations, including the incompressible Navier-Stokes equations, high-dimensional heat equations, highdimensional transport equations, and dispersive equations of higher-order derivatives.
What problem does this paper attempt to address?
### Problems Addressed by the Paper
This paper aims to address the efficient solution of high-dimensional time-dependent partial differential equations (PDEs). Specifically, the authors propose Partial Evolution Tensor Neural Networks (pETNNs), a novel framework capable of handling high-dimensional time-dependent PDEs with high accuracy. Traditional numerical methods such as finite difference, finite element, and spectral methods experience an exponential increase in computational cost when dealing with high-dimensional problems, known as the "curse of dimensionality," which severely limits their application in high-dimensional systems.
### Main Contributions
1. **Enhanced Computational Framework**: Combining the Tensor Neural Networks (TNNs) framework with evolution parameter approximation, enabling it to handle time-dependent PDEs while maintaining high accuracy.
2. **Innovative Update Strategy**: Introducing a partial parameter update strategy that optimizes the allocation of computational resources, improving system efficiency.
3. **A Posteriori Error Estimation**: Proposing a posteriori error bounds for pETNNs based on dynamic parameter approximation theory, applicable to various spatial discretization schemes.
4. **Empirical Validation**: Numerical experiments demonstrate that pETNNs can efficiently solve complex PDEs, including the incompressible Navier-Stokes equations, high-dimensional heat equations, high-dimensional transport equations, and dispersion equations with higher-order derivatives, without sacrificing accuracy while reducing computational costs.
### Method Overview
- **Tensor Neural Networks (TNNs)**: TNNs decompose high-dimensional integral problems into a series of one-dimensional integral problems, using Gaussian quadrature formulas to replace Monte Carlo integration, significantly improving solution accuracy.
- **Partial Evolution Tensor Neural Networks (pETNNs)**: Building on TNNs, introducing a time-dependent parameter evolution mechanism, achieving efficient solutions for time-dependent PDEs through a partial parameter update strategy.
- **Parameter Update Strategy**: Employing a predictor-corrector method (i.e., modified Euler method) to numerically integrate the parameter evolution equations, ensuring long-term stability of the model.
- **A Posteriori Error Estimation**: Proposing a posteriori error bounds for pETNNs through perturbation analysis of linear ODE systems, providing theoretical support for the model's accuracy.
### Numerical Experiments
- **Comparative Experiments**: Comparing the performance of pETNNs with other methods (such as PINNs, EDNNs, and ETNNs) through examples like the transport equation, incompressible Navier-Stokes equations, high-dimensional heat equations, high-dimensional transport equations, and simplified Korteweg-de Vries equations.
- **Parameter Update Strategy**: Testing both full parameter update and partial parameter update strategies, with results showing that the partial parameter update strategy significantly reduces computation time while maintaining high accuracy.
### Conclusion
By combining Tensor Neural Networks and evolution parameter approximation, pETNNs successfully address the efficient solution of high-dimensional time-dependent PDEs. This method not only outperforms traditional methods in terms of accuracy but also demonstrates significant advantages in computational efficiency, making it particularly suitable for handling high-dimensional complex systems.