覃一澜,马嘉禹,付海洋,等. 基于JEC-FDTD等效循环神经网络的电磁建模和等离子体参数反演[J]. 电波科学学报,2024,39(3):1-9. DOI: 10.12265/j.cjors.2023217
      引用本文: 覃一澜,马嘉禹,付海洋,等. 基于JEC-FDTD等效循环神经网络的电磁建模和等离子体参数反演[J]. 电波科学学报,2024,39(3):1-9. DOI: 10.12265/j.cjors.2023217
      QIN Y L, MA J Y, FU H Y, et al. Electromagnetic modeling and plasma parameters inversion based on JEC-FDTD equivalent recurrent neural network[J]. Chinese journal of radio science,2024,39(3):1-9. (in Chinese). DOI: 10.12265/j.cjors.2023217
      Citation: QIN Y L, MA J Y, FU H Y, et al. Electromagnetic modeling and plasma parameters inversion based on JEC-FDTD equivalent recurrent neural network[J]. Chinese journal of radio science,2024,39(3):1-9. (in Chinese). DOI: 10.12265/j.cjors.2023217

      基于JEC-FDTD等效循环神经网络的电磁建模和等离子体参数反演

      Electromagnetic modeling and plasma parameters inversion based on JEC-FDTD equivalent recurrent neural network

      • 摘要: 磁化等离子体中的电磁波传播是重要的研究课题,针对特定场景下的电磁等离子耦合问题,进行有效且准确的方程建模与参数求解具有极强的研究价值和挑战性,这是探究电磁波与等离子体复杂非线性相互作用机制的关键。文中设计了一种可用于电磁等离子体正逆向建模的循环神经网络(recurrent neural network, RNN),该网络正向传播过程等价于任意磁倾角情况下的电流密度卷积时域有限差分 (current density convolution finite-difference time-domain, JEC-FDTD)方法,因此可以求解给定的电磁建模问题,并易于大规模并行计算。通过构建前向可微模拟过程,JEC-FDTD方法可以使用自动微分技术准确且高效地计算梯度,然后通过训练网络来解决反问题。因此,该方法可以有效利用观测到的时域散射场信号反演重要的等离子体参数。JEC-FDTD方法和RNN相结合,形成了较强的协同效应,使得模型具有可解释性和高效的计算效率,受益于深度学习提供的优化策略和专用硬件支持,可以适用于不同仿真场景下的电磁建模和等离子体参数反演。

         

        Abstract: Electromagnetic wave propagation in magnetized plasma stands as a significant focal point in research. The effective formulation and precise resolution of equations to model electromagnetic plasma coupling within specific scenarios carry profound research significance and pose noteworthy challenges. These undertakings are pivotal in delving into the intricate mechanisms governing the nonlinear interaction between electromagnetic waves and plasma. This paper introduces a novel approach employing recurrent neural networks (RNNs) for both forward and backward modeling of electromagnetic plasma. The forward propagation process emulates the current density convolution finite difference time-domain method (JEC-FDTD), accommodating arbitrary magnetic tilts. Consequently, it not only resolves targeted electromagnetic modeling issues but also offers the advantage of efficient computation through parallel processing. By establishing a forward-differentiable simulation process, our methodology adeptly leverages automatic differentiation techniques to precisely compute gradients. This enables us to effectively address inverse problems through network training. As a result, our approach harnesses observed time-domain scattered field signals to successfully deduce pivotal plasma parameters. Taking advantage of the optimization strategy and dedicated hardware support provided by deep learning, the method can be applied to electromagnetic modeling and plasma parameter inversion in various simulation scenarios.

         

      /

      返回文章
      返回