Please use this identifier to cite or link to this item: https://dspace.ncfu.ru/handle/20.500.12258/26574
Title: Solving Poisson Equation by Physics-Informed Neural Network with Natural Gradient Descent with Momentum
Authors: Abdulkadirov, R. I.
Абдулкадиров, Р. И.
Lyakhov, P. A.
Ляхов, П. А.
Nagornov, N. N.
Нагорнов, Н. Н.
Keywords: Finite element method;Poisson equation;Fisher matrix;Neural networks;Natural gradient descent
Issue Date: 2023
Citation: Abdulkadirov, R.I., Lyakhov, P.A., Nagornov, N.N. Solving Poisson Equation by Physics-Informed Neural Network with Natural Gradient Descent with Momentum // Proceedings of the Seminar on Signal Processing, SoSP 2023. - 2023. - pp. 3-6. - DOI: 10.1109/IEEECONF60473.2023.10366065
Series/Report no.: Proceedings of the Seminar on Signal Processing, SoSP 2023
Abstract: In this paper, we present a solution of the first boundary value problem of the Poisson equation using physics-informed neural networks, which optimize the loss function by natural gradient descent with Dirichlet distribution. This algorithm proved his ability in optimization of multidimensional functions with high accuracy, taking little number of iterations and avoiding local minimums. This fact means, that proposed algorithm will minimize loss function as well. Also, we obtain the exact and numerical solutions, implied due to separation of variable (Fourier method) and finite element method, respectively. The main goal of this paper is to rise the accuracy of solution of the Poisson equation in a short time. We will demonstrate the error estimate of solutions received by finite element method and physics-informed neural network, equipped with adaptive moment estimates and natural gradient descent with momentum and Nesterov condition.
URI: http://hdl.handle.net/20.500.12258/26574
Appears in Collections:Статьи, проиндексированные в SCOPUS, WOS

Files in This Item:
File SizeFormat 
scopusresults 2957 .pdf
  Restricted Access
132.27 kBAdobe PDFView/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.