EFnet: an added value of multi-model simulation Mikhail Sofiev Finnish Meteorological Institute

Content Introduction: why the ensemble modelling Examples of single- and multi-model ensembles EFnet: ensemble forecasting and model development Conclusions

Ensemble modelling: why? Atmospheric processes are stochastic The smaller scale and the shorter averaging the higher uncertainty small-scale processes, as well as some chemical chains of reactions can be chaotic by nature Deterministic models work poor at small scales, with short averages and complicated chemical chains. Reason is NOT (well, not only) model weaknesses but rather stochastic nature of the atmosphere Right form of question: probability terms Ways to answer the probabilistic questions make probabilistic models (what about physics?) run ensembles of existing deterministic model(s)

Types of ensembles Single-model multi-setup ensemble One deterministic model Input forcing, initial and/or boundary conditions are perturbed in a “reasonable” way or taken from several sources Each perturbed set of data is computed in a normal way Output datasets are considered as realizations of a stochastic process Example: ECMWF ensemble weather forecast (operational !) Multi-model ensemble Several deterministic (and/or other) models are used Each model uses own input datasets and/or common set(s) Example: EU FP5 ENSEMBLE project, NKS MetNet network, EMEP Pb-1996 model inter-comparison

Single-model ensemble: ECMWF Source: www.ecmwf.int

Multi-model ensemble: NKS MetNet

Multi-model ensemble: EU-ENSEMBLE project Source: Galmarini, 2004

Multi-model ensemble: EMEP Pb model inter-comparison Source: Sofiev et al., 1996

Multi-model debugging Source term are the same for both models Meteorological data are the same but: Models used own meteo pre-processors Source: Potempski, 2005

EFnet: forecasting and model development Validation of the model results Operational Quality Assurance QA: simple, basic statistics, provides the basic quality scores for individual models Scientific assessment of the model quality: detailed speciation, precursor analysis, detailed statistics annual model inter-comparison exercises during high O3, PM seasons Statistical correction of the model forecasts Based on past model quality scores – both operational and scientific QA Model-specific some models may already have it depends on the model individual quality score

EFnet: forecasting and model development (2) Ensemble forecast (feasibility study) Straightforward generation of statistically-sounding ensemble is far beyond the reach of current computers Multi-model results allow approaching an uncertainty-disclosing problem in air quality prediction problem Predictability of ozone and PM levels Points of chaos, multi-track developments, etc Output: ensemble average (if possible) + uncertainty range

Conclusions Multi-model ensemble is a brand-new tool in air quality assessment First ensembles were applied in emergency modelling deterministic models work very poorly established co-operation, need for mutual backup comparatively straightforward (cheap) application Outcome from the ensemble usage MUCH better stability than individual models show uncertainty ranges, areas/times of instability, errors or model failures often agrees with measurements better than individual models EFnet ensemble approach Build and understand the multi-model ensemble for O3 and PM If successful, generate the ensemble air quality forecast