Machine-learned boundary conditions for the optimal absorption of charged particles in particle-in-cell simulations
ORAL
Abstract
When running numerical simulations of open systems, solvers in the simulation domain must be complemented by appropriate absorbing boundary conditions (BCs) to minimize the impact of a finite box size. In Particle-in-Cell simulations, Perfectly Matched Layers (PMLs) are a standard method for absorbing electromagnetic radiation that would escape the domain. However, developing analogous BCs for the absorption of charged particles remains a long-standing problem.
Recent progress has been made in extending the PML method to the special case of absorbing particles crossing the simulation boundary at normal incidence [1], but this scheme continues to produce significant spurious radiation for charged particles at arbitrary angles of incidence. In this work, we frame the design of absorbing BCs for charged particles as an optimization problem, and solve it via gradient-based methods using a differentiable Maxwell solver written in PyTorch [2]. We find new absorbing BCs that outperform previous methods both at normal and oblique incidence. Our results demonstrate the potential of leveraging differentiable solvers to design new numerical algorithms.
[1] R. Lehe et al., ArXiv:2201.09084 [Physics] (2022).
[2] A. Paszke et al., Adv. in Neur. Info. Proc. Sys. 32, 2019, pp. 8024–8035.
Recent progress has been made in extending the PML method to the special case of absorbing particles crossing the simulation boundary at normal incidence [1], but this scheme continues to produce significant spurious radiation for charged particles at arbitrary angles of incidence. In this work, we frame the design of absorbing BCs for charged particles as an optimization problem, and solve it via gradient-based methods using a differentiable Maxwell solver written in PyTorch [2]. We find new absorbing BCs that outperform previous methods both at normal and oblique incidence. Our results demonstrate the potential of leveraging differentiable solvers to design new numerical algorithms.
[1] R. Lehe et al., ArXiv:2201.09084 [Physics] (2022).
[2] A. Paszke et al., Adv. in Neur. Info. Proc. Sys. 32, 2019, pp. 8024–8035.
–
Presenters
-
Mark Almanza
University of California, Los Angeles, UCLA, Department of Physics and Astronomy, Los Angeles, CA, USA
Authors
-
Mark Almanza
University of California, Los Angeles, UCLA, Department of Physics and Astronomy, Los Angeles, CA, USA
-
Amadou Diallo
Lawrence Berkeley National Laboratory, Berkeley, CA, USA
-
Edoardo Zoni
Lawrence Berkeley National Laboratory, Lawrence Berkeley National Laboratory, Berkeley, CA, USA
-
Remi Lehe
Lawrence Berkeley National Laboratory, Lawrence Berkeley National Laboratory, Berkeley, CA, USA
-
Paulo Alves
University of California, Los Angeles, UCLA, Department of Physics and Astronomy, Los Angeles, CA, USA, UCLA