SPECTRAL NEURAL NETWORK PRECONDITIONER FOR THE DISCRETE DIPOLE APPROXIMATION
Section: OPTICAL INVESTIGATION OF ATMOSPHERE
Authors:
1.
Institut optiki atmosfery im. V.E. Zueva SO RAN
2.
National Research Tomsk State University
3.
Institut optiki atmosfery im. V.E. Zueva SO RAN
4.
Institut optiki atmosfery im. V.E. Zueva SO RAN
Nacional'nyy issledovatel'skiy Tomskiy gosudarstvennyy universitet
Nacional'nyy issledovatel'skiy Tomskiy gosudarstvennyy universitet
Type:
Сonference article
Pages:
from 213
to 217
Published:
17.08.2026
Subject area:
UDC 504.3
Language:
Russian
Keywords:
discrete dipole method, light scattering, neural network preconditioner, iterative methods, fast Fourier transform
Funding:
Calculations using the discrete dipole method were carried out within the framework of the state assignment of the Institute of Atmospheric Optics SB RAS, and the neural network was trained within the framework of the Russian Science Foundation project No. 25-17-00087, https://rscf.ru/project/25-17-00087/.
Abstract:
This work presents a neural network preconditioner for the discrete dipole approximation that substantially accelerates electromagnetic scattering computations for arbitrarily shaped particles. Standard iterative solvers lose efficiency at high refractive indices and fine discretizations. The proposed architecture encodes particle geometry and optical parameters, then predicts spectral corrections to the inverse interaction matrix in the Fourier domain. Integration of the module into the ADDA code achieves up to 48 times speedup in iteration count and ensures convergence in previously divergent cases. Experiments confirm robust performance for refractive indices with real parts exceeding 2.0. The approach is effective within the training distribution but requires retraining for each new configuration and currently does not generalize to unseen particle shapes.
This work presents a neural network preconditioner for the discrete dipole approximation that substantially accelerates electromagnetic scattering computations for arbitrarily shaped particles. Standard iterative solvers lose efficiency at high refractive indices and fine discretizations. The proposed architecture encodes particle geometry and optical parameters, then predicts spectral corrections to the inverse interaction matrix in the Fourier domain. Integration of the module into the ADDA code achieves up to 48 times speedup in iteration count and ensures convergence in previously divergent cases. Experiments confirm robust performance for refractive indices with real parts exceeding 2.0. The approach is effective within the training distribution but requires retraining for each new configuration and currently does not generalize to unseen particle shapes.
Keywords:
discrete dipole method, light scattering, neural network preconditioner, iterative methods, fast Fourier transform
discrete dipole method, light scattering, neural network preconditioner, iterative methods, fast Fourier transform
1. Yurkin M.A., Hoekstra A.G. The DDA code ADDA: Capabilities and known limitations // JQSRT. 2011. V.112. P.2234–2247.
2. ADDA. https://github.com/adda-team/adda
3. Van der Vorst H.A. Bi-CGSTAB // SIAM J. Sci. Stat. Comput. 1992. V.13. P.631–644.
4. Hausner P. et al. Neural Incomplete Factorization // arXiv:2305.16368. 2023.
