ncof development workshop 2008
DESCRIPTION
NCOF Development Workshop 2008. Assessments of Ecosystem Models using Assimilation Techniques. John Hemmings, Peter Challenor, Ian Robinson & Tom Anderson. What is the “Ecosystem Model” in Ecosystem Model Assessment ?. Ocean Biogeochemical General Circulation Model. Free-running model - PowerPoint PPT PresentationTRANSCRIPT
NCOF Development Workshop 2008
Assessments of Ecosystem Models using Assimilation Techniques
John Hemmings, Peter Challenor, Ian Robinson & Tom AndersonJohn Hemmings, Peter Challenor, Ian Robinson & Tom Anderson
What is the “Ecosystem Model” inEcosystem Model Assessment ?
• Free-running modelFree-running model• Assimilation system (sequential D.A.)Assimilation system (sequential D.A.)
Ocean Biogeochemical General Circulation ModelOcean Biogeochemical General Circulation Model
Ecosystem Sub-modelEcosystem Sub-model
• Fixed parameter modelFixed parameter model• Model structure and formulationModel structure and formulation
Outline
• The Calibration Process (Inverse D.A. Scheme)The Calibration Process (Inverse D.A. Scheme)• Allowing for UncertaintyAllowing for Uncertainty• Assessment of D.A. Scheme and Model Assessment of D.A. Scheme and Model • Combining Data from Different LocationsCombining Data from Different Locations
• Sequential Assimilation of Ocean ColourSequential Assimilation of Ocean Colour• Improving Forecasts and HindcastsImproving Forecasts and Hindcasts
The Calibration Process
ECO.MODEL
OPTIMIZER
COSTFUNC.
SimulatedObs.
MisfitCost
Calibration Obs.
BoundaryConditions
ForcingInitial
Conditions
FreeParameters
ScienceOutput
Sensitivity Analysis
ValidationObs.
Allowing for Uncertainty
The Misfit Formulation
Estimate 2SIM by:
1) Characterizing uncertainty in IC, physical forcing & boundary fluxes2) Propagating through model by ensemble runs
Misfit =
(xSIM - xOBS)2
2DEP
2DEP = 2
OBS + 2SIM
For a given parameter set, 2SIM is uncertainty due to
IC, physical forcing & boundary fluxes
Allowing for Uncertainty
External Input Data for 1-D Simulations
• Biogeochemical tracer profiles Biogeochemical tracer profiles BBi i (z, member)(z, member)
Initial conditions:
Forcing data:
• Sea-surface PAR Sea-surface PAR I (t, member)I (t, member)• Sea-surface salinity Sea-surface salinity S (t, member)S (t, member)• Mixed layer depth Mixed layer depth M (t, member)M (t, member)• Temperature Temperature T (z, t, member)T (z, t, member)• Vertical diffusion coefficient Vertical diffusion coefficient k (z, t, member)k (z, t, member)• Vertical velocity Vertical velocity w (z ,t, member)w (z ,t, member)
Boundary fluxes:
• Horizontal biogeochemical tracer fluxes Horizontal biogeochemical tracer fluxes HHi i (z, t, member)(z, t, member)
Allowing for Uncertainty
Marine Model Optimization Test-bed (MarMOT)
INPUT ITEMS (1 or more instances of each)
physical forcing
run options:ecosystem model, time-step, misfit spec. …
initial conditions
boundary conditions
observations
fixed parametersMODEL SPECIFIC
N SITES
N SITES
N SITES
N SITES
misfit costother validation stats.
model outputM CASES
free parameters (posterior)
MODEL SPECIFIC
free parameters (prior)
MODEL SPECIFIC
case table
Generic Function Analyzer
Model Evaluator
(1-D)Optimizermisfit
cost
Assessment CriteriaAssimilation Scheme & Calibration Data Set
• Fit to data from non-calibration yearsFit to data from non-calibration years- better than prior parameter set - better than prior parameter set
TWIN EXPERIMENTS REAL-WORLD EXPERIMENTS
• True solution known• Can test parameter recovery
• Ecosystem is real
• Idealized scenario may be unrepresentative
• Uncertainty in IC, forcing, horizontal fluxes and observations affects validation misfit
+
-
• No. of parameters constrained (with acceptable repeatability)No. of parameters constrained (with acceptable repeatability)
Assessment CriteriaEcosystem Model
Calibrated Model:
• Fit to data from non-calibration yearsFit to data from non-calibration years- better than cal. data climatology - better than cal. data climatology
Model Structure and Formulation:
• Fit to data from non-calibration yearsFit to data from non-calibration years- better than alternative model with same cal. data better than alternative model with same cal. data
Limitation: optimal calibration not possible for complex models
Ecosystem Model Assessment
An Example Model Comparison Experiment
OG99 NPZD:OG99 NPZD: Oschlies and Garçon (1999) Oschlies and Garçon (1999)HadOCC NPZD:HadOCC NPZD: Hadley Centre Ocean Carbon Cycle Model, Hadley Centre Ocean Carbon Cycle Model,Palmer and Totterdell (2001) - modifiedPalmer and Totterdell (2001) - modified
Thanks to Ben Ward & Andrew Yool for providing OCCAM output at BATS
Combining Data from Different Locations
Identifying Calibration Provinces
NERC Data Assimilation Thematic Programme
Zero-D NPZ model fit to daily Zero-D NPZ model fit to daily chlorophyll + winter nitrate at chlorophyll + winter nitrate at calibration stationscalibration stations
Split-domain calibration method Split-domain calibration method (Hemmings, Srokosz, Challenor & (Hemmings, Srokosz, Challenor & Fasham, 2004):Fasham, 2004):
identifies optimal geographic identifies optimal geographic ranges for single parameter sets ranges for single parameter sets by selecting promising stations to by selecting promising stations to aggregateaggregate
Final provinces chosen by Final provinces chosen by misfit cost at validation stationsmisfit cost at validation stations
Sequential Assimilation of Ocean ColourCASIX Chlorophyll Assimilation Scheme in FOAM-HadOCC
3D analysis
2D analysis of log(Chl)
2D analysis of P
ΔN
ΔP
ΔZ
ΔD
Δalk
ΔDIC
Model forecast
N:Chl
Observations
• Aim: improve air-sea COAim: improve air-sea CO22 flux by improving surface DIC and alkalinity, hence pCO flux by improving surface DIC and alkalinity, hence pCO22
• 2-D analysis of log2-D analysis of log1010(Chl) uses FOAM analysis correction scheme (as for SST)(Chl) uses FOAM analysis correction scheme (as for SST)• Surface phytoplankton increments derived using model nitrogen:chl (dynamic)Surface phytoplankton increments derived using model nitrogen:chl (dynamic)• Other variables adjusted by a new material balancing scheme (Hemmings, Barciela & Other variables adjusted by a new material balancing scheme (Hemmings, Barciela & Bell, 2008) Bell, 2008)
Rosa Barciela, Matt Martin, Mike Bell, Adrian Hines (Met Office)John Hemmings (NOCS)
DAILY ANALYSIS CYCLE
Sequential Assimilation of Ocean ColourMaterial Balancing Scheme for Nitrogen and Carbon
• Surface phytoplankton increment Surface phytoplankton increment given as inputgiven as input
• Relative increments to other nitrogen Relative increments to other nitrogen pools depend on the likely pools depend on the likely contributions to phytoplankton error contributions to phytoplankton error from growth and loss from growth and loss
• Nitrogen conserved at each grid Nitrogen conserved at each grid point (if possible)point (if possible)
• DIC increment conserves carbonDIC increment conserves carbon
• Sub-surface scheme prevents Sub-surface scheme prevents formation of unrealistic sub-surface formation of unrealistic sub-surface minima in DINminima in DIN
Sequential Assimilation of Ocean ColourEvaluation of Material Balancing in 1-D Twin Experiments
Free runFree run
Assimilating Chl & PAssimilating Chl & P
Assimilating Chl onlyAssimilating Chl only
60ºN
40ºN
50ºN
30ºN
Sequential Assimilation of Ocean Colour3-D Evaluation of Chlorophyll Assimilation Scheme
Biogeochemical errors due to excessive vertical transport of nutrients not corrected by Biogeochemical errors due to excessive vertical transport of nutrients not corrected by chlorophyll assimilation (intentionally)chlorophyll assimilation (intentionally)
TWIN EXPERIMENTSTWIN EXPERIMENTS REAL-WORLD EXPERIMENTSREAL-WORLD EXPERIMENTS
Surface ChlorophyllSurface Chlorophyll
Un-assimilated VariablesUn-assimilated Variables
Need biogeochemical Need biogeochemical balancing scheme when balancing scheme when assimilating T&S profilesassimilating T&S profiles
Impact of Physical D.A.(link to MARQUEST)
DIN
Chlorophyll
physics DA on
DA off
physics DA onDA off
??
Improving Forecasts and Hindcasts: the Role of Parameter OptimizationA Non-identical Twin Experiment
Truth:Truth: HadOCC HadOCCEcosystem Model:Ecosystem Model: Simplified HadOCC with 4 free parameters Simplified HadOCC with 4 free parametersCalibration data:Calibration data: Chlorophyll (daily), DIN & pCO Chlorophyll (daily), DIN & pCO22 (monthly) (monthly)
Improving Forecasts and Hindcasts: the Role of Parameter OptimizationSequential Chlorophyll Assimilation Results
TRUTH
ORIGINAL
ORIGINAL + CHL D.A.
OPTIMIZED
OPTIMIZED + CHL D.A.
Surface Chlorophyll
Surface Phytoplankton
Surface DIN
Surface pCO2
Improving Forecasts and HindcastsApplication of Different Assimilation Methods
Sequential Data Assimilation
• Improve hindcast stateImprove hindcast state• Improve initial conditions for short-term forecastsImprove initial conditions for short-term forecasts
Parameter Optimization (Inverse D.A. Methods)
• Improve long-term forecastImprove long-term forecast• Improve performance of sequential schemesImprove performance of sequential schemes