Member Profile

Nathan Eizenberg

PhD student

School of Earth Sciences
University of Melbourne

neizenberg@student.unimelb.edu.au

Biography

Nathan is a PhD student at the University of Melbourne investigating ensemble data assimilation techniques with Professor Craig Bishop. He has an MSc in mathematical modelling and scientific computing from the University of Oxford (2015), a BSc from Monash University (2013) and last year completed an honours degree in meteorology at the University of Melbourne (2020). He has worked as a research scientist at the Bureau of Meteorology in data assimilation, primarily on the project to produce the first atmospheric regional reanalysis for Australia (BARRA). As well as research, Nathan has an interest in science communication and advocacy and has been involved in Pint of Science and has performed at Melbourne's Laborastory.

THESIS: Local ensemble tangent linear model for 4D variational data assimilation schemes

Assimilation of observations into weather models is a key process for systematically reducing forecast error. Particularly when forecast conditions are far from the climatological mean, i.e extreme events, other bias correcting methods which rely on good statistical precedence are unhelpful. Ensemble forecast provide likelihood information about forecast weather and these rely on good perturbed initial conditions often from a DA scheme. Data assimilation also has a role in the calibration and improvement of climate and weather models. Climate model forecasts initial conditions and parameterisations can be tuned to minimise the deviation from observations over a spin-up period. We can also use model error covariances produced by data assimilation to improve parameterisation schemes or understand how climate models from different research centres are correlated to each other. Reanalyses are also products of data assimilation which provide dynamically consistent representation of past climate in model space. However, state of the art data assimilation algorithms are often expensive to develop and maintain such that often less sophisticated schemes are used. For example, the 4DVar scheme relies on having a seperate, linearised version of the nonlinear model, called the tangent linear model (TLM) and adjoint, to produce model initial conditions which also minimise error from observations recorded hours away from the forecast start time. Producing and maintaining a good TLM requires expert knowledge and comprehensive understanding of the nonlinear models which are being constantly updated. They are also dependant on model scale and poorly represent physical processes which are nonlinear in nature like precipitation and cloud formation. This research project will investigate methods of using ensembles of forward model trajectories to produce a TLM and adjoint without the separate linearised codebase. This could enable the use of sophisticated variational data assimilation schemes without the historical cost of maintaining them and without the need to recode the TLM every time there is a change in the nonlinear model. As well as assisting operational weather centres such an improvement in the modelling and assimilation machinery would hopefully strengthen modelling research by making the state-of-the-art accessible for all projects.