Data Assimilation in Geosciences

Transcription

1 Data Assimilation in Geosciences Alberto Carrassi The Nordic Centre of Excellence for ensemble-based data assimilation Laurent Bertino (Lead), Alberto Carrassi (Co-Lead), Colin Grudzien (PD), Patrick Raanes (PD), Matthias Rabatel (PD), Abhishek Shah (PhD), Maxime Tondeur (MsC) and Colin Guider (PhD visitor)

2 Data Assimilation - Outlook Model Simulations Observations Data Assimilation best combines model and observations and brings synergy An on-going rapid expansion from Weather Science (NWP) into Climate Science/Geophysics in general: Oceanography Climate Prediction Climate Assessment Hydrology Geology Climatology Detection & Attribution and many more beyond geosciences NWP Geophysics

3 Problem Statement environmental system observations forecast analysis Ø The problem is in principle solved using a probabilistic framework Ø The quantity of interest are the probability density functions, PDF model truth time Ø The PDFs are evolved in time and updated at analysis times using Bayes s rules

4 DA methods in geoscience: David vs Goliath Climate models have huge dimension: 10 9 variables Environmental observations constitute a massive dataset: 10 7 daily obs Climate phenomena are intrinsically nonlinear and chaotic Nonlinear observation EnKF NERSC world-class milestone Evensen, 1994 Present 1990s Ensemble Kalman Filter/ Smoother 1980s 4D variational assimilation 1970s 3D variational assimilation Optimal Interpolation q q Nonlinear Obs q Gaussian q Multivariate q q Nonlinear optimisation q Time dimension q Implicit Zlow dependent error description q Bauer et al., 2015 q Explicit Zlow dependent error description q No need for tangent and adjoint models q Scalable & Parallelizable

5 DA for geosciences The role of NERSC Evensen, Bertino, Sakov, Counillon, Raanes, Xie, Lisæter, Carrassi and many other TOPAZ SST reanalysis for 2007 Jiping Xie Ø Worldwide leadership in the theory of ensemble DA methods Stochastic and square-root formulations Parameter estimation Anamorphosis for non-gaussian variables Ø Parallel operational development Real-time DA for ocean and sea-ice (TOPAZ) Ocean and sea-ice reanalysis (1990-present) Climate predictions (NorESM)

6 Nowadays challenges for geophysical DA Where are we going? ü Coupled Data Assimilation ü Fully Bayesian method (nonlinear and non-gaussian) ü Lagrangian Data Assimilation ü Computationally efzicient DA Big Data Prediction*time*horizon* hours& weeks& years& Fully&Bayesian& nonlinear&da& Big#Data## Nonlinear# DA# Particle*Filter* Turbulence& Convection* Coupled#DA# EnVar* Weather*&*Ocean** Prediction** Gaussian& nonlinear&da& EnKF*/*4DVar* Seasonal;to;Decadal* Prediction** ENSO* KF*/*3DVar* Mul;variate&Gaussian& quasi.linear&da& DA#complexity# Fully.coupled&DA& Weakly.coupled&DA& Uncoupled&DA& 10.3& Km& 10 Model*grid*size* 2& Km& Carrassi et al., 2016

7 Coupled Data Assimilation Ø Improve the forecast of coupled phenomena (hurricanes, coastal weather, ENSO, MJO) Ø Use Earth System Simulators as the unizied modeling instrument across all timescales from days to decades Ø Better exploit the new generation of Earth observations (Argo, SMOS...) Ø Produce coupled reanalysis Key ScientiZic Issue Ø Presence of multiple scale of motions Ø Information propagation across model compartment Ø DifZicult to compute cross-covariance Ø NERSC pioneer in the coupled DA for ocean and sea-ice (Lisæter et al. 2003) Ø NERSC DA group is conducting basic research on coupled DA and applied-oriented investigation in collaboration with the Climate Dynamics and Prediction group.

8 EfSicient Bayesian Data Assimilation Ø EnKF/4DVar fails in highly nonlinear/non-gaussian situations Ø Nonlinear Bayesian Particle Filter are required. Key ScientiZic Issue Ø Curse of Dimensionality Ø Big Data Problem (model/obs 10 9 /10 7 ) Ø Not computational power alone Unknown&True& PDF& EnKF&& Smart &PF& Sample'where' PDF'is'large'!''' Ø NERSC DA group is actively studying advanced formulations of Particle Filter and EnKF to deal with nonlinearity Ø The main idea is to study PF which incorporates model dynamic's features in its design (see Raanes and Grudzien poster)

9 Lagrangian Data Assimilation DA for the Sea Ice Assimilation of data instruments or for model grid that are not stationary is known as Lagrangian DA Lagrangian models Lagrangian observations in the Arctic Credit: P.Rampal Lagrangian Data Assimilation Ø Ø Ø Ø Mesh varies in time Remeshing procedures which does not conserve the total number of elements New formulation of DA methods are required A collaboration between DA and Sea-Ice modeling groups is studying extensions of the EnKF for time-evolving and remeshed grid (see Matthias Rabatel poster)