Next Generation Multiscale Adaptive Mesh Atmospheric Modelling, Rapid Response and Data Assimilation

Transcription

1 Next Generation Multiscale Adaptive Mesh Atmospheric Modelling, Rapid Response and Data Assimilation Fangxin Fang, Jeff Gomes 29, August, 2017

2 AMCG website (

3 Open Source Model Software for Multiphysics Problems Unstructured FEM Meshes Anisotropic Adaptive Mesh technology Fluidity User-friendly GUI Python interface to calculate diagnostic fields, to set prescribed fields and userdefined boundary conditions

4 Predictive and Uncertainty Model Framework Uncertainty quantification Analysis, real-time updating and representation of uncertainties. Data science laboratory Atmospheric Modelling Multi-scale predicative modelling with multiphysical Fluidity Optimal design, risk assessment, decision and policy-making processes risk Data assimilation Optimisation of uncertainties Minimisation of misfit between observational data and numerical results HPC Reduced order modelling Emergency rapid response Ocean, flooding multifluid flow modelling Adaptive targeted observations Optimal experimental design Optimal sensor locations

5 Computational Multi-Fluid Flow Dynamic Model Porous media model embedded in Fluidity are based on: new family of FE element-pairs (P n DG-P m and P n DG-P m DG) and numerical formulation (overlapping CVFEM) that ensures high-order accuracy on the solution fields (i.e., pressure, velocity, saturation, temperature etc) The model has been used on transport of contaminant in subsurface media (mining spillage), oil and gas production, nuclear waste repository etc;

6 Atmospheric Environmental Model: Critical bridge between human activities and environmental change (IAP) Severely Polluted Day PM 2.5 ~470 µg m -3 Polluted Day PM 2.5 ~200 µg m -3 Clean Day PM 2.5 ~15 µg m -3 Beijing, Olympic site

7 Atmospheric Environmental Model: Critical bridge between human activities and environmental change (IAP) Environmental issues Agriculture and Ecosystem Human Activities (urbanization, motorization and industrialization, emissions) CO 2, CO, NO 2, SO 2, O 3, aerosols, etc Climate issues Human Health Effects Atmospheric dynamical transport Atmospheric chemical transformation Radiative forcing (scatting,absor bing, cloud) Climate Change Regional air pollution Global air pollution Natural emissions(carbon, nitrogen) Ocean System Land, Water, Snows Deposition (Acid rain, etc) Terrestrial Ecosystems Air pollution is largely influenced by human activities, and can be reduced by human.

8 Urban Air Pollution Modelling (MAGIC, supported by Castle, London

9 Air Pollutant Modelling (IAP-ICL, supported by NSFC/ EPSRC) SO 2 released from over 100 power plants (over Beijing and 55 cities)

10 Chemical modelling (IAP-ICL, Zheng etc.) (NO and NO2 released from over 100 power plants) Continuous formation/consumption of NOx and Ozone over 5 days.

11 3D chemical modelling (IAP-ICL, Zheng etc.) (NO and NO2 released from over 100 power plants) 1 day

12 _ _ _18 Dust storm (IAP-ICL, Zheng etc.) _04

13

14 3D Atmospheric Modelling: Cyclone 3D Simulation (IAP-ICL) Potential Temperature Rain Vapor Water Velocity Vector Adaptive mesh Cloud

15 Tohoku event from 2011 (Visiting PhD student from Japan)

16 Natural Disaster: Simulating Flooding Denmark (R. Hu etc. supported by the EU PEARL project) Fig. 11. Situation of Study Area in Greve, Municipality of Denmark, see Soledad (2014)

17 Natural Disaster: Simulating Flooding Denmark (R. Hu etc. supported by the EU PEARL project) Surface Surface with mesh

18 Predictive and Uncertainty Model Framework Uncertainty quantification Analysis, real-time updating and representation of uncertainties. Atmospheric Modelling Multi-scale predicative modelling with multiphysical Fluidity Data assimilation Optimisation of uncertainties Minimisation of misfit between observational data and numerical results Data science laboratory Optimal design, risk assessment, decision and policy-making processes risk HPC Reduced order modelling Emergency rapid response Ocean, flooding multifluid flow modelling Adaptive targeted observations Optimal experimental design Optimal sensor locations

19 DA methods: Optimal interpolation; Nudging; 3D-Var; 4D-Var (Adjoint); Ensemble KF Data Assimilation (DA) Motivation for DA: To improve the predictability of numerical models; Uncertainty sensitivity analysis; Optimisation of uncertainties in models; Goal-based error measure and mesh adaptivity; Design optimisation; Adaptive observation (Optimisation of sensors locations).

20 Work from Institute of Atmospheric and Physics, China PM 2.5 before DA PM 2.5 after DA Differences Sources: CNEMC IAP

21 DA: Optimal sensor location

22 Reduced Order Modelling Navier-Stokes Equations: Discretising equations (1) and (2) Full discretised system (high dimensional): Projecting onto the reduced space using SVD/POD Wring the reduced order model in a general form The function f is represented by deep learning

23 Reduced Order Modelling (POD) and Deep Learning (Xiao etc) Top panel: Full modelling

24 Rapid modelling: Air flow (Elephant Castle London) (Xiao etc. supported by MAGIC EPSRC) Left: reduced order modelling; right: high fidelity modelling CPU time: seconds (Reduced order model); 3 hours (10 cores, Full fidelity model)

25 Future work Reginal Model Global Model Fluidity NAQPM WRF DA Adaptive-mesh air quality model Gas chemistry Adaptive-mesh weather forecast Microphysics HPC Future work Liquid chemistry Cumulus parameterization Future work PM chemistry Surface Layer Planetary Boundary Layer, Radiation

26 Thanks

27 Rapid modelling: Flow past two buildings Top: reduced order modelling; bottom: high fidelity modelling