Abstract : Mullti-physics problems typically have essential dynamics. Numerical models are often hindered by difficulties in fully
capturing the relevant physics. Nowadays, more attention has been given to data science approaches. A hybrid AI
and multiscale physical modeling approach could be the optimal way to provide a dynamic understanding of the
governing equations. Data-driven modeling results may find some patterns which are not expected from physical
modeling. This session aims at exploring the challenges of physical and data-driven modeling for real-time
prediction and applications to geosciences, addressing uncertainty quantification, data assimilation, high-performance
computing, machine learning, numerical methods, and reduced order modeling.
Organizer(s) : Haroldo Fraga de Campos Velho, Fangxin Fang
02562 (1/2) : 2C @G703 [Chair: Haroldo F. de Campos Velho]
[04758] Data driven modelling and EnKF for spatial-temporal forecasting: Ozone and PM forecasting in China
Format : Talk at Waseda University
Author(s) :
Fangxin Fang (Imperial College London)
Meiling Cheng (Imperial College London)
Shengjuan Cai (Imperial College London)
Christopher Pain (Imperial College London)
Yanghua Wang (Imperial College London)
Michael I Navon (Florida State University)
Jiang Zhu (Institute of Atmospheric Physics, Chinese Academy of Sciences)
Jie Zhu (Institute of Urban Environment)
Jinxi Li (Institute of Atmospheric Physics, Chinese Academy of Sciences)
Xiaofei Wu (Chengdu University of Information Technology)
Abstract : Spatiotemporal forecasting involves generating temporal forecasts for system state variables across spatial regions. Data-driven methods, such as Convolutional Long Short-Term Memory (ConvLSTM) and deep convolutional generative adversarial network (DCGAN), are effective in capturing both spatial and temporal correlations. To further improve the predictive accuracy, the data assimilation EnKF is introduced to data driven modelling. Here, the performance of the data driven models has been demonstrated in hourly and daily spatiotemporal pollutant forecasting in China. The results have been compared to monitoring measurements and physical modelling results.
[05103] Prediction of Swirling Fluid Flow Pattern in a River
Abstract : Swirling of water in a river occurs when rocks, holes, obstacles, or sudden changes in the river channel obstruct the flow of the water. The swirl is characterized by turbulent parameters (velocities and Reynolds stresses) and corresponding factors associated with turbulence. The main objective of this study is to obtain the governing equations for swirling flow and to predict the flow pattern. The proposed theoretical models show very good agreement with experimental data.
[05293] A fast, high resolution pluvial flood model for risk assessment and real-time flood prediction
Format : Online Talk on Zoom
Author(s) :
Steven Cocke (Florida State University)
Dong-Wook Shin (Florida State University)
Abstract : A high resolution, computationally efficient pluvial flood model has been developed to provide flash flood inundation estimates due to heavy precipitation events. The need for a computationally fast model is critical for estimating flood risk, where a large number of flood scenarios are needed to obtain a reliable probability distribution of flood depths and extents, as well as for real-time prediction where sufficient advance warning must be given to the public.
[05285] Machine learning for data assimilation and predictability to the atmospheric models
Format : Talk at Waseda University
Author(s) :
Haroldo Fraga de Campos Velho (INPE: National Institute for Space Research)
Rosangela Cintra (INPE: National Institute for Space Research)
Steven Cocke (FSU: Florida State University)
Vinicius Albuquerque Almeida (UFRJ: Federal University of Rio de Janeiro)
Juliana Aparecida Anochi (INPE: National Institute for Space Research)
Vinicius Monego (INPE: National Institute for Space Research)
Abstract : Data assimilation is one of the most important challenges for the computational effort of the operational centers for weather and climate predictions. In this talk, the use of machine learning approaches will be shown for numerical weather models. Techniques for data assimilation for global and regional models are addressed by artificial neural networks. The analysis computed by self-configuring a supervised neural network for the COAPS-FSU global model is designed to emulate the Local Ensemble Transform Kalman filter. A deep learning neural network is applied to the WRF-NCAR regional model as a new method for data assimilation, where the 3D-Var scheme is employed as a reference to the machine learning approach. Our numerical experiments show a significant reduction in the CPU-time to calculate the analysis maintaining the precision of the forecasting for both models. Finally, another important issue is to evaluate how good is the prediction, in other words, how we can calculate the forecasting confidence interval. The standard procedure to compute the confidence interval is to apply the ensemble prediction. A novelty approach to estimate the prediction uncertainty quantification is addressed by using machine learning algorithms: neural networks, and decision tree formulations.
02562 (2/2) : 2D @G703 [Chair: Prof. Fangxin Fang]
[04677] Aadaptive mesh atmospheric model development
Format : Online Talk on Zoom
Author(s) :
Jinxi Li (Institute of Atmospheric Physics, Chinese Academy of Sciences)
Fangxin Fang (Imperial College London)
Pu Gan (Chengdu University of Information Technology)
Christopher Pain (Imperial College London)
Xiaofei Wu (Chengdu University of Information Technology)
Zifa Wang (Institute of Atmospheric Physics, Chinese Academy of Sciences)
Jie Zheng (Institute of Urban Environment, Chinese Academy of Sciences)
Jiang Zhu (Institute of Atmospheric Physics, Chinese Academy of Sciences)
Abstract : This study presents the development of a three-dimensional unstructured adaptive finite-element model (Fluidity-Atmosphere) for atmospheric research. To improve the computational efficiency, a LSTM-based three-dimensional unstructured mesh generator is proposed to predict the evolution of the adaptive mesh. To evaluate the performance of adaptive meshes and physical parameterisations in Fluidity-Atmosphere, a series of idealized test cases have been setup and the unstructured tetrahedral meshes are adapted automatically with the specified fields in time and space.
