Scholarlink

<p>Shrublands and grasslands, which constitute approximately 70% of Australia’s vegetation, play a critical role in global wildfire-prone regions. To advance the understanding of grass fire spread, a three-dimensional, physics-based fire model provides valuable insights into fire dynamics. However, such models are computationally intensive and time-consuming. To address these challenges, we constructed an extensive numerical database comprising 64,000 high-fidelity wildfire simulation cases and implemented a Long Short-Term Memory neural network architecture. The model demonstrates strong predictive performance, achieving a coefficient of determination (<i>R</i><sup>2</sup>) of 0.96 on training data, indicating excellent agreement with the physics-based simulation outputs. By utilizing coordinates from five reference points to predict fire front movement, this approach offers a novel method for analysing fire dynamics in homogeneous fuel beds with an average deviation of less than 2.5%. Combining the strengths of physics-based modelling and deep learning, our research enhances fire spread prediction accuracy of over 95% while significantly reducing computational demands. Future efforts will focus on refining the model, expanding the dataset, and incorporating additional variables to improve predictive capabilities and operational applicability.</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

Integrating deep learning with physics-based model for predicting grassfire spread

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月18日

查看全文
  • Rahul Wadhwani,
  • Xiaoning Zhang,
  • Yizhou Li,
  • Duncan Sutherland,
  • Khalid Moinuddin,
  • Xinyan Huang

摘要

Shrublands and grasslands, which constitute approximately 70% of Australia’s vegetation, play a critical role in global wildfire-prone regions. To advance the understanding of grass fire spread, a three-dimensional, physics-based fire model provides valuable insights into fire dynamics. However, such models are computationally intensive and time-consuming. To address these challenges, we constructed an extensive numerical database comprising 64,000 high-fidelity wildfire simulation cases and implemented a Long Short-Term Memory neural network architecture. The model demonstrates strong predictive performance, achieving a coefficient of determination (R2) of 0.96 on training data, indicating excellent agreement with the physics-based simulation outputs. By utilizing coordinates from five reference points to predict fire front movement, this approach offers a novel method for analysing fire dynamics in homogeneous fuel beds with an average deviation of less than 2.5%. Combining the strengths of physics-based modelling and deep learning, our research enhances fire spread prediction accuracy of over 95% while significantly reducing computational demands. Future efforts will focus on refining the model, expanding the dataset, and incorporating additional variables to improve predictive capabilities and operational applicability.