Scholarlink

Abstract <p>The paper examines a methodological approach to simulating nonlinear atmospheric dynamics. The approach implies constructing a surrogate (replacement) model of a physical object (process) based on machine learning. For illustrative purposes, the surrogate model is built for the conceptual model of the coupled ocean–atmosphere system, in which the atmospheric component is represented by a low-dimensional nonlinear dynamic system, and the harmonic oscillator model is used as the oceanic component. The surrogate model is based on unidirectional and bidirectional long short-term memory neural networks (LSTM and BiLSTM, respectively). Nine LSTMs and one BiLSTM, whose structures were determined experimentally, are analyzed to evaluate their ability to predict complex dynamics and chaotic regime in the examined model on time intervals from 5 to 10 days. The best forecast accuracy was obtained using the BiLSTM.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Surrogate Modeling Methodology for Nonlinear Atmospheric Dynamics: From Conceptual Model to Neural Networks

  • 期刊论文

  • 发表日期：

    2025年2月1日

查看全文
  • S. A. Soldatenko,
  • Ya. I. Angudovich

摘要

Abstract

The paper examines a methodological approach to simulating nonlinear atmospheric dynamics. The approach implies constructing a surrogate (replacement) model of a physical object (process) based on machine learning. For illustrative purposes, the surrogate model is built for the conceptual model of the coupled ocean–atmosphere system, in which the atmospheric component is represented by a low-dimensional nonlinear dynamic system, and the harmonic oscillator model is used as the oceanic component. The surrogate model is based on unidirectional and bidirectional long short-term memory neural networks (LSTM and BiLSTM, respectively). Nine LSTMs and one BiLSTM, whose structures were determined experimentally, are analyzed to evaluate their ability to predict complex dynamics and chaotic regime in the examined model on time intervals from 5 to 10 days. The best forecast accuracy was obtained using the BiLSTM.