Learning Unsplit-field-based PML for the FDTD Method by Deep Differentiable Forest
This is an incremental improvement for computational electromagnetics researchers, offering a more efficient boundary condition method.
The paper tackles the problem of reducing computational complexity in finite-difference time-domain (FDTD) simulations by proposing a deep differentiable forest (DDF) model to replace the traditional perfectly matched layer (PML) absorbing boundary condition, achieving a reduction to a one-cell thickness boundary layer with satisfactory numerical accuracy.
Alternative unsplit-filed-based absorbing boundary condition (ABC) computation approach for the finite-difference time-domain (FDTD) is efficiently proposed based on the deep differentiable forest. The deep differentiable forest (DDF) model is introduced to replace the conventional perfectly matched layer (PML) ABC during the computation process of FDTD. The field component data on the interface of traditional PML are adopted to train the DDF-based PML model. DDF has the advantages of both trees and neural networks. Its tree structure is easy to use and explain for the numerical PML data. It has full differentiability like neural networks. DDF could be trained by powerful techniques from deep learning. So compared to the traditional PML implementation, the proposed method can greatly reduce the size of FDTD physical domain and the calculation complexity of FDTD due to the novel model which only involves the one-cell thickness of boundary layer. Numerical simulations have been carried out to benchmark the performance of the proposed approach. Numerical results illustrate that the proposed method can not only easily replace the traditional PML, but also be integrated into the FDTD computation process with satisfactory numerical accuracy and compatibility to the FDTD.