Chemical Data Assimilation in Support of Chemical Weather

Chemical Data Assimilation in Support of Chemical Weather Forecasts Greg Carmichael, Adrian Sandu, Dacian Daescu, Tianfeng Chai, John Seinfeld, Tad Anderson, Peter Hess, Dacian Daescu Data Assimilation

Chemical Data Assimilation in Support of Chemical Weather Forecasts Outline Ø Motivation Ø Current State of Forward Models Ø Data Assimilation Framework (4 d. Var) – Issues Ø Preliminary Results Ø Future Directions

Models are an Integral Part of Atmospheric Chemistry Studies • Flight planning • Provide 4 -Dimensional context of the observations • Facilitate the integration of the different measurement platforms • Evaluate processes (e. g. , role of biomass burning, heterogeneous chemistry…. ) • Evaluate emission estimates (bottom-up as well as top-down) • Emission control strategies testing • Air quality forecasting

TRACE-P/Ace. Asia/ITCT-2 K 1 EXECUTION Stratospheric intrusions Satellite data in near-real time: MOPITT TOMS SEAWIFS AVHRR LIS FLIGHT PLANNING Long-range transport from Europe, N. America, Africa Boundary layer chemical/aerosol processing ASIAN OUTFLOW 3 D chemical model forecasts: -x - GEOS-CHy. EM - CFORS -z PACIFIC ASIA Emissions -Fossil fuel -Biomass burning -Biosphere, dust PACIFIC

Forward Models Are becoming More Comprehensive MOZART Global Chemical Transport Model Influence Functions Emission Biases/ Inversion Mesoscale Meteorological Model (RAMS or MM 5) Anthropogenic & biomass burning Emissions Meteorological Dependent Emissions (biogenic, dust, sea salt) TOMS O 3 STEM Tracer Model (classified tracers for regional and emission types) Airmasses and their age & intensity Analysis STEM Prediction Model with on-line TUV & SCAPE Chemistry & Transport Analysis STEM Data. Assimilation Model Observations

Fight Planning: Frontal outflow of biomass burning plumes E of Hong Kong 100 ppb Biomass burning CO forecast Observed CO –Sacshe et al. Observed aerosol potassium - Weber et al. Longitude

Predictability – as Measured by Correlation Coefficients Met Parameters are Best < 1 km Performance decreases with altitude

Model vs. Observations + Modeled O 3 vs. Measured O 3 • Cost functional measures the modelobservation gap. • Goal: produce an optimal state of the atmosphere using: Ø Model information consistent with physics/chemistry Ø Measurement information consistent with reality

Development of a General Computational Framework for the Optimal Integration of Atmospheric Chemical Transport Models and Measurements Using Adjoints (NSF ITR/AP&IM 0205198 – Started Fall 2002) A collaboration between: Greg Carmichael (Dept. of Chem. Eng. , U. Iowa) Adrian Sandu (Dept. of Comp. Sci. , Virginia Tech. ) John Seinfeld (Dept. Chem. Eng. , Cal. Tech. ) Tad Anderson (Dept. Atmos. Sci. , U. Washington) Peter Hess (Atmos. Chem. , NCAR) Dacian Daescu (Dept. Math, Portland State) http: //atmos. cgrer. uiowa. edu/people/tchai/

Basic Idea of 4 D-Var • Define a cost functional • Derive adjoint of tangent linear model Where adjoint variables are the sensitivities of the cost functional with respect to state variables (concentrations), i. e. Useful by themselves !! • Update Initial conditions using the gradients

Assimilation Results ØAssimilate O 3/NO 2 with O 3/NO 2 observations in the window [0, 6] GMT, March 01, 2001; ØTwin experiments framework; ØFull 3 D simulation with SAPRC chemical mechanism. O 3

CO-assimilation

Observation Frequency vs Number of Species O 3 & NO 2 O 3 - only

Recovery of O 3 and NO 2 is Different WHY? NO 2 O 3

20% 1% Most of the grid points values are recovered within in 1%; but some locations the error is > 20%. Additional details see Chai’s paper on Thursday Assimilation requires better algorithms (with known error behavior)

Overview of Research in Data Assimilation for Chemical Models. Solid lines represent current capabilities. Dotted lines represent new analysis capabilities that arise through the assimil. of chemical data. Ensemble methods

Chemical Assimilation and Big-Iron “BIGMAC”@VT üRanked 3 rd with measured performance = 10 Tflop/s. üA Pentium class cluster with 16 -24 processors has ~ 50 Gflop/sec. üOn such a cluster we run parallel STEM (Trace. P): 1 hour simulation time / 5 minutes cpu time üOn the terrascale machine we can run in parallel an ensemble of 200 simulations for the same simulation / cpu time ratio.

Assimilation of Aerosol Dynamics Gradient Methods Data Frequency Recovery of Initial Distribution • Theoretical framework enables the solution of coupled coagulation and growth with minimal number of size bins; • Piecewise polynomial discretizations; • Adjoint/assimilation system built

We plan to test some of these developments in an operational setting this summer as part of a large field experiment.

We are Developing General Software Tools to Facilitate the Close Integration of Measurements and Models The framework will provide tools for: 1) construction of the adjoint model; 2) handling large datasets; 3) checkpointing support; 4) optimization; 5) analysis of results; 6) remote access to data and computational resources. http: //atmos. cgrer. uiowa. edu/people/tchai/ Adjoints being developed for MOZART, plans for WRF-Chem

Chemical Data Assimilation: The Future? ü Feasible & necessary. ü Just the beginning— more ? ? s than answers – but we have test beds! ü Huge implications for measurement systems and models. ü Need to grow the community. TWO-SCENARIO FORECAST

http: //www. wmo. ch/web/arep/gaw/urban. html

Air Quality Forecasting Research Elements Summary of USWRP Air Quality Forecasting Workshop April 29 - May 1, 2003 Houston, TX