Scholarlink

Abstract <p>In this study, we test the capability of machine learning (ML) methods to approximate the characteristics of passive scalar transport in turbulent flows within the planetary boundary layer (PBL). Training and validation datasets were generated from a suite of large-eddy simulation (LES) experiments, ensuring physically consistent reference solutions. Several ML models were implemented and evaluated, including random forests, gradient boosting, and fully connected neural networks (FCNNs). Their predictive performance was benchmarked against conventional approaches, namely the empirical Briggs plume model and linear regression. The results show that state-of-the-art ML algorithms can significantly improve the approximation of plume statistics compared to baseline methods. Among the tested models, the FCNN provided the highest accuracy, achieving a coefficient of determination (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^{2}\)</EquationSource> <!--BPhysMGU2570309Gerasimov-m1--> </InlineEquation>) of 0.93 for the lateral and vertical standard deviations of scalar concentration and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(R^{2}=0.70\)</EquationSource> <!--BPhysMGU2570309Gerasimov-m2--> </InlineEquation> for the vertical displacement of the plume centerline. The corresponding root mean square errors (RMSE) for key plume parameters were also small. A feature-importance analysis further confirmed that the models captured physically meaningful dependencies: the dominant predictors were geometric factors such as downwind distance and source height, as well as thermodynamic drivers including the surface heat flux and the temperature gradient above the PBL.</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

Modeling Turbulent Transport of Passive Scalars in the Planetary Boundary Layer Using Large Eddy Simulation and Machine Learning

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • I. A. Gerasimov,
  • M. A. Krinitskiy,
  • E. V. Mortikov

摘要

Abstract

In this study, we test the capability of machine learning (ML) methods to approximate the characteristics of passive scalar transport in turbulent flows within the planetary boundary layer (PBL). Training and validation datasets were generated from a suite of large-eddy simulation (LES) experiments, ensuring physically consistent reference solutions. Several ML models were implemented and evaluated, including random forests, gradient boosting, and fully connected neural networks (FCNNs). Their predictive performance was benchmarked against conventional approaches, namely the empirical Briggs plume model and linear regression. The results show that state-of-the-art ML algorithms can significantly improve the approximation of plume statistics compared to baseline methods. Among the tested models, the FCNN provided the highest accuracy, achieving a coefficient of determination ( \(R^{2}\) ) of 0.93 for the lateral and vertical standard deviations of scalar concentration and \(R^{2}=0.70\) for the vertical displacement of the plume centerline. The corresponding root mean square errors (RMSE) for key plume parameters were also small. A feature-importance analysis further confirmed that the models captured physically meaningful dependencies: the dominant predictors were geometric factors such as downwind distance and source height, as well as thermodynamic drivers including the surface heat flux and the temperature gradient above the PBL.