First SAEMC-IAI, STIC-AmSud Workshop. 5 June 2007

Transcription

1 First SAEMC-IAI, STIC-AmSud Workshop Meryem Ahmed de Biasi, Irène Korsakissok, Vivien Mallet Lin Wu 5 June 2007

2 Outline

3 Images Greek Mythology, cyclops in Odyssey Why this name? Poly : multiple phemus : speech

4 Multiple Multiple Models Scales : from local scale to continental scale Formulations : Gaussian, Eulerian,... Multiple Pollutants Passive, radionuclides Photochemistry Aerosols Persistent organic pollutants, heavy metals,... Multiple Inputs From meteorological models : MM5, ECMWF Ground data : GLCF, USGS

5 Multiple Methods Sequential Variational Inverse modeling (parameter estimation) Ensemble Forecast Multimodels Monte Carlo Models combinations ( superensembles,... ) Models Coupling Feedbacks Impact

6 Constraints Perennial Code System maintenance Scalable, integration of new developments Open Availability, distribution Development or contributions from other teams Field Context From research to operational use

7 Overall structure base Files processing libraries (Atmo, Seldon,...) Input data processing Computes physical fields Files I n p u t D R I V E R Numerical integration (Polair3D) Physics O u t p u t Statistics (AtmoPy) Libraries with physical parameterizations (Atmo)

8 Libraries Talos : operations on configuration files management of dates Seldon : input-output operations Atmo : extension of Seldon to atmospheric sciences, physical parameters AtmoPy : postprocessing Preprocessing Meteorological fields from ECMWF or MM5 Ground data : L Use Cover, Emission, Deposition Boundary conditions from Mozart 2 or Inca for gaseous species Gocart for aerosols Gaussian models.

9 Models Gaussian Models Stationary Gaussian model puff Gaussian model Gas aerosols Several parameterizations to compute the dispersion Eulerian Models Castor (clone of the gas version of Chimere) Polair3D (passive, chemical, aerosol adjoint [for gas] versions) Modules for transport, chemistry aerosols Postprocessing Comparison to observations Water diagnosis in a plume Visualization

10 Programming Choices Main language : C++ Efficient for scientific computing Advanced object design Inheritance, genericity Management of complex objects Exceptions Widely used perennial

11 Programming Choices Complementary Language : Python Dynamic, interactive Visualization Scripts Increasingly used in scientific computing

12 Programming Choices Historical Language : Fortran 77 Policy Automatic differentiation Continuity Calls from C++ To avoid dealing with too many languages To use primarily languages with strong potential productivity

13 Application Main Reference Main article for : "Technical Note : The air quality modeling system ", V. Mallet, D. Quélo, B. Sportisse, M. Ahmed de Biasi, É. Debry, I. Korsakissok, L. Wu, Y. Roustan, K. Sartelet, M. Tombette H. Foudhil, Atmos. Chem. Phys. Discuss., 7 (3), 2007

14 Partners Main Organizations 1 ENPC 2 EDF Impact studies of thermal power plants 3 INRIA data assimilation French partners 1 IRSN (Nuclear safety) Radionuclides 2 INERIS (industrial risks) Local scale, ozone forecasting, liquid water content diagnosis 3 CETE (urban equipment) Evaluation of Urban mobility plan 4 DGA (Defense) Local scale, network evaluation International collaborations 1 CRIEPI in Japan Sensitivity studies over Europe, Paris Tokyo 2 EDF Polska Impact studies, air quality modeling 3 IER Stuttgart, European projects Health impact evaluation 4 CMM in Chile, STIC-AMSUD project Air quality forecasting, data assimilation

15 Some Aerosols at Regional Continental Scales Polair3D hosts an aerosol model (SIREAM). Validations over Europe Asia, over Paris, Marseille Lille. Sensitivity studies over Europe, Paris Tokyo. - M. Tombette B. Sportisse. Aerosol modeling at regional scale : Model-to-data comparison sensitivity analysis over Greater Paris. Atmos. Env., In press - K. Sartelet, E. Debry, K. Fahey, M. Tombette, Y. Roustan, B. Sportisse. Simulation of aerosols gas phase species over Europe with the POLYPHEMUS system. Part I : model-to-data comparison for year Atmos. Env., doi : /j.atmosenv K. Sartelet, H. Hayami, B. Sportisse. Dominant aerosol processes during high-pollution episodes over Greater Tokyo. J. Geophys. Res., Accepted for publication

16 Some Impact Studies at Continental Scale Sensitivity study of ozone concentrations with respect to emissions Evaluation of air quality in Polska Evaluation of the emissions from French thermal power plants for the year 2005 lat PM2.5 mean lon Vivien Mallet Bruno Sportisse. A comprehensive study of ozone sensitivity with respect to emissions over Europe with a chemistry-transport model. J. Geophys. Res., 110(D22), 2005

17 Some Local scale models Comparison of sigma parameterizations Network design for defense applications Plume in Grid model Better modeling of major point source emissions Application to passive tracers (ETEX, Chernobyl) - Irène Korsakissok. Performance evaluation of Gaussian models with Prairie Grass experiment. Technical Report , CEREA, 2007

18 Some Air Quality Modeling over Lille Assessment of health impact of the Urban Mobility Plan (CETE) Coupling with emission traffic model, indicators of health impact Inverse modeling of emissions with 4D-var second-order sensitivity - Denis Quélo, Vivien Mallet, Bruno Sportisse. Inverse modeling of NO x emissions at regional scale over Northern France. Preliminary investigation of the second-order sensitivity. J. Geophys. Res., 110(D24310), D. Quélo, R. Lagache, B. Sportisse. Etude de l impact qualité de l air des scénarios PDUs sur Lille à l aide du modèle de Chimie-Transport Polair3D. Technical Report , CEREA, Rapport PREDIT - R. Lagache, C. Declercq, D. Quélo, B. Sportisse, P. Palmier, B. Quetelard, F. Haziak. Evaluation du PDU de Lille-Métropole sur le trafic, les concentrations de polluants atmosphériques et la mortalité. In Actes de la 15ième conférence sur les transports et la pollution de l air, 2006

19 Some Air Quality Ensemble Forecast Estimation of the uncertainties in the output Model sequential agregation with weights learned over the past Ozone forecasting over Europe (Prév air operational platform). - Vivien Mallet Bruno Sportisse. Ensemble-based air quality forecasts : a multi-model approach applied to ozone. J. Geophys. Res., 111(D18) :18302, Vivien Mallet Bruno Sportisse. Uncertainty in a chemistry-transport model due to physical parameterizations numerical approximations : an ensemble approach applied to ozone modeling. J. Geophys. Res., 111(D01302), 2006

20 Some Dispersion of Radionuclides is the support of the new operational forecast system at the Emergency Center of IRSN Simulation of the Chernobyl accident, sensitivity studies Inverse modeling applied to simulations of the ETEX exercise Algeciras accident [Bq.m^{-3}] Measurements [Bq.m^{-3}]

21 Some Chernobyl Accidental Release, 25 April-5 May 1986 run, Forecast Emergency Center IRSN/CEREA T13:03:00 Chernobyl 1E-01 1E+00 1E+01 1E+02 - D. Quélo, M. Krysta, M. Bocquet, O. Isnard, Y. Minier, B. Sportisse. Validation of the POLYPHEMUS system : the ETEX, Chernobyl Algeciras cases. Atmos. Env., doi : /j.atmosenv B. Sportisse. A review of parameterizations for modeling dry deposition scavenging of radionuclides. Atmos. Env., (41) : , 2007

22 Why? To improve the results (analysis) using several sources of information (model, observations). How? assimilation involves three component : A model (currently only possible with Polair3D model gaseous species) Observations An assimilation algorithm which combines both (implemented as a driver in )

23 Notations x : model state vector. x t : true state. x b : background estimates. x a : analysis. y : observation vector. ǫ b : x b x t background errors. ǫ a : x a x t analysis errors. H : Observation operator B (also noted P f ) : (ǫ b ǫ b )(ǫ b ǫ b ) T background error covariance matrix. A (also noted P a ) : (ǫ a ǫ a )(ǫ a ǫ a ) T analysis error covariance matrix. ǫ o : y H(x t ) observation errors. R : (ǫ o ǫ o )(ǫ o ǫ o ) T observation error covariance matrix.

24 Optimal Interpolation Principle Combination of background background departure. Formulae x a = x b + K(y H(x b )), K = BH T (HBH T + R) 1. Study B : using Balgovind correlation function. f(r) = ( 1 + L) r ( ) exp r L v b Perturbs ICs (upper). Biased model (lower). Ozone concentration at grid point [15, 33] Ozone concentration at grid point [15, 33] 160 reference Perturbation Optimal interpolation Hours 200 reference Biased model simulation Optimal interpolation Hours

25 Ensemble Kalman filter (EnKF) Principle Time-dependent error covariance matrices. Ensemble statistics are used to approximate P f P a at each time. Formulae Initialization : given initial pdf p(x t 0 ), an ensemble of r members are generated romly, P a 0 = 1 r 1 {x a,i 0,, i = 1,..., r} x a 0 = r 1 r x a,i 0 i=1 r ( ) ( x a,i 0 x a 0 i=1 x a,i 0 x a 0 ) T

26 EnKF Formulae Forecast formula : x f,i k P f k = 1 r 1 where x f k Analysis formula : = M k 1[x a,i k 1 ] + ηi k 1 r ( x f,i k x f k i=1 )( x f,i k x f k is the mean of ensemble {xf,i k x a,i k P a k = 1 r 1 = x f,i k + K ( ) k y i k H k[x f,i k ] r i=1 r x a k = 1 r i=1 ( x a,i k x a k x a,i k )( x a,i k x a k ) T, i = 1,.., r} ) T

27 Reduced rank square root Kalman filter : RRSQRT Heemink et al Principle Also approximate error covariance matrices, but using a low-rank representation LL T. L is the Mode Matrix whose q columns represent the dominant directions of the forecast. In addition, square root of Q (the error model) is approximated using ensemble statistics. Formulae Initialization : x a 0, La 0 = [la,1 0,...,la,q 0 ]

28 RRSQRT Formulae Forecast formula : x f k = M k 1 k(x a k 1 ) l f,i k = 1 ǫ {M k 1 k(x a k 1 +ǫla,i k 1 ) M k 1 k(x a k 1 )}, ǫ = 1 L f k = [lf,1 k,..., lf,q k, Q 1 2 k 1 ], L f k = Πf L k f k Analysis formula : P f k = Lf k Lf,T k K k = P f k HT k (H kp f k HT k + R k) 1, x a k = xf k + K k(y k H(x f k )), L a k = [(I K kh k )L f k, K kr 1 2 k ], L a k = Πa k L a k

29 Four-Dimensional Variational : 4D-Var A cost function J(x) that deals with a set of observations : 1 2 (x xb ) T B 1 (x x b ) + 1 }{{} 2 J b N {}}{ (y k H k (x k )) T R 1 k (y k H k (x k )) k=0 } {{ } J o where the assimilation window is 0 N, x k = M 0 k (x) = M k M k 1...M 1 x The gradient is calculated by the backward integration of adjoint model x N = 0 For k = N,..., 1, calculates x k 1 = M T k ( x k H T k d ) k, where d k = R 1 k (y k H k (x k )) x 0 := x 0 H T 0 (d 0) gives the gradient of J o with respect to x J oi

30 Results RMSE reference OI EnKF 4DVar RRSQRT Concentration µg m reference observation OI EnKF 4DVar RRSQRT Hours Root mean square errors Observation numbers Results at Monton station.

31 Current Developers 1 Meryem Ahmed de Biasi (INRIA) diffusion 2 Édouard Debry (ENPC) aerosols 3 Karine Kata-Sartelet (ENPC) aerosols 4 Irène Korsakissok (ENPC) local, plume-in-grid 5 Vivien Mallet (ENPC) ensemble 6 Denis Quélo (IRSN) passive 7 Yelva Roustan (ENPC) impact 8 Bruno Sportisse (ENPC) aerosols 9 Marilyne Tombette (ENPC) aerosols 10 Lin Wu (INRIA) assimilation

32 Documentations for users User s guide (150 pages) Scientific documentation for Atmo Reference documentation for AtmoPy Documentations for developers Guide reference documentation for Seldon Reference documentation for Atmo Guide reference documentation for Talos Examples Test cases (Eulerian Gaussian) Practical sessions (primarily for courses at ENSTA ENPC)

33 A Graphical User Interface

34

35 Models (mostly implemented, but not yet included in ) Secondary organic aerosols Modal aerosol module (MAM) Heavy metals, mercury Persistent organic pollutants (POP) Passive hemisphere model Next Steps Parallelization Lagrangian particle model Drivers for assimilation coupling