Data assimilation in Earth system models combines high-dimensional, coupled, nonlinear models with large volumes of in situ and remotely sensed observational data. The dynamics and observations are nonlinear, the distributions are non-Gaussian, and the cost of simulation is high. The goal of the minisymposium is to provide a forum for this diverse group to discuss and share ideas for advancing the science of DA in climate modeling or any of its components (e.g. atmosphere, ocean, ice sheets, land models, or sea ice). Possible topics of interest include coupled data assimilation; strategies for estimating and mitigating model errors; strategies for addressing strong nonlinearities and non-Gaussianity; multiscale, multilevel, or multifidelity methods; and machine learning methods for data assimilation.
14:00
EnKF-FAQ
Patrick N. Raanes | NORCE | Norway
Show details
Author:
Patrick N. Raanes | NORCE | Norway
The ensemble Kalman filter (EnKF) is a popular data assimilation technique. This talk answers some questions that practitioners of the EnKF may ask, namely:
- Regarding its linearizations:
* What exactly are they?
* Why does this rarely get mentioned?
* How does it relate to analytic derivatives?
- Regarding its covariances:
* Why are the estimates normalized by (N-1)?
* Why do we prefer the Kalman gain "form" in the EnKF?
- Regarding nonlinear models:
* Why does it create sampling error?
* Why does it cause divergence?
14:30
- CANCELED - A Particle Filter-based Adaptive Inflation Scheme for the Ensemble Kalman Filter
Ibrahim Hoteit | KAUST | Saudi Arabia
Show details
Authors:
Boujemaa Ait-El-Fquih | KAUST | Saudi Arabia
Ibrahim Hoteit | KAUST | Saudi Arabia
This talk is devoted to the problem of loss of ensemble variance in the ensemble Kalman filter (EnKF). Adaptive inflation methods are solutions mostly based on a Bayesian approach, which considers the inflation factor as a random variable with a given prior probability distribution, and then combines it with the inflation likelihood through Bayes' rule to obtain its posterior distribution. In this work, we introduce a numerical implementation of this generic Bayesian approach that uses a particle filter (PF) to compute a Monte Carlo approximation of the inflation posterior distribution. To alleviate the sample attrition issue, the proposed PF employs an artificial dynamical model for the inflation factor based on the well-known smoothing-kernel West and Liu model. The positivity constraint on the inflation factor is further imposed through an inverse-Gamma transition density, whose parameters suggest analytical expressions. The resulting PF-EnKF scheme is straightforward to implement, and can use different numbers of particles in its EnKF and PF components. Numerical experiments that are conducted with the Lorenz-96 model to demonstrate the effectiveness of the proposed method will be presented.
15:00
A hybridized EnKF and smoothed-observations particle filter
Ian Grooms | University of Colorado Boulder | United States
Show details
Authors:
Ian Grooms | University of Colorado Boulder | United States
Gregor Robinson | University of Colorado | United States
Ensemble Kalman Filters in many forms and permutations are very successful methods for large-scale data assimilation. The EnKF performance is limited by two kinds of errors: Sampling errors and errors associated with the implicit assumption of Gaussianity. Both types of errors can be mitigated in a variety of ways. We present a method for mitigating the errors associated with non-Gaussianity in the specific case where the prior (forecast) distribution is manifestly non-Gaussian while the posterior is close to Gaussian. Such situations arise when the observations (data) are plentiful, and have small, Gaussian-distributed errors. Our method is based on splitting the likelihood L(x)^a L(x)^(1-a), 0 < a < 1. A particle filter is applied first, followed by an EnKF. The particle filter produces an intermediate distribution that is presented to the EnKF as a prior. If the splitting factor "a" is chosen appropriately, this intermediate distribution is much closer to Gaussian than the prior. Particle filters perform poorly in high dimensions, so we limit the effective dimensionality by localizing in length scale: Observation errors at small scales are modeled as being much higher than they are in reality. Done carefully, this improves the particle filter performance. The method is tested on a multiscale Lorenz-96 model, and we show that the hybrid performs better than the pure EnKF. Our example uses an ensemble square root version of the EnKF with a mean-preserving random rotation.
15:30
- CANCELED - Impact of ocean observation systems on ocean analyses and subseasonal forecasts
Aneesh Subramanian | University of Colorado Boulder | United States
Show details
Author:
Aneesh Subramanian | University of Colorado Boulder | United States
We evaluate the relative merits of different ocean observation systems (moored buoys, Argo, satellite, XBTs and others) by their impact on ocean analyses and subseasonal forecast skill. Several ocean analyses were performed where different ocean observation platforms were withheld from the assimilation in addition to one ocean analysis where all observations were assimilated. We then use these ocean analyses products for initializing a set of subseasonal forecasts to evaluate the impact of different ocean analyses states on the forecast skill. We use the European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble prediction system for the twenty-year sub-seasonal hindcast experiments. Results from these hindcast experiments will be presented to highlight changes in the ocean analyses states and their impact on the forecast skill of the MJO, monsoon intraseasonal oscillations, as well as global temperature and precipitation.
Coupled air-sea interaction processes relevant to intraseasonal variability (e.g. the MJO, MISO) in the earth’s climate system are inadequately represented in coupled models. New efforts in observations, process understanding and translation into weather and climate models are necessary for improvements in simulation and prediction of the intraseasonal variability and associated weather events. We discuss the merits of different ocean observation platforms in this context and also future observation and model improvement pathways.