Skip to content


    Skip to content

     

    COMPUTATIONAL FLUID DYNAMICS | Contents | Next

    Forecasts for the Weather of the Earth's Oceans

    FEATURED
    Carl Wunsch, Cecil and Ida Green Professor of Physical Oceanography
    John Marshall, Professor of Atmospheric and Oceanic Sciences
    Center for Global Change Science, Department of Earth, Atmospheric and Planetary Sciences, Massachusetts Institute of Technology

    We need weather forecasting because we live on land. We want to know if a hurricane or tornado is near, if drought will kill the crops, or if the tomatoes will freeze. We don't live in the oceans, and the creatures that do have their own ways of coping with changes in their environment. So do we need to know the state of the oceans? We do, scientists say, if we want to be able to predict the climate, which is the weather over long periods of time. And a group of ocean scientists at MIT has been making use of NPACI resources to build a system to forecast ocean "weather."

    In contrast to short-term weather patterns, the global climate--over months, seasons, years, decades, and longer--is a combination of the behavior of the terrestrial biosphere, the atmosphere, and the oceans. And the oceans determine the outcomes more and more strongly the longer the time frame.

    COLLECTING OCEAN WEATHER DATA

    FORWARD AND ADJOINT MODELS

    THE EXPERIMENT


    Global Circulation ModelFigure 1: Global Circulation Model

    Results from the global optimization of the MIT GCM, surface forcing fields, and one year of TOPEX/POSEIDON data. Top: Time mean velocity estimate at 60-meter depth and the surface elevation in centimeters. Middle: Potential temperature and flow vectors at 610 meters (level 10). Bottom: Same as above, except at 2450 m (level 15).

    COLLECTING OCEAN WEATHER DATA

    Meteorological forecast models take data from a vast array of sources --satellites, ground stations, and midair observations--and use those data in solving the equations governing the circulation of the atmosphere, producing weather predictions. "Data on the circulation of the oceans is harder to obtain, so it is tougher to make good estimations of ocean 'weather,'" said Carl Wunsch. Wunsch, the Cecil and Ida Green Professor of Physical Oceanography, and John Marshall, professor of atmospheric and oceanic sciences, lead the team of ocean scientists at the Massachusetts Institute of Technology (MIT).

    "We are using our primary ocean general circulation model, or OGCM, constructed by my colleague John Marshall and his group," Wunsch said, "together with data from many different sources in an effort to obtain an understanding of the absolute, time-varying, large-scale circulation of the oceans, and its impact on climate." The work is part of a Climate Model Initiative at MIT begun in 1994, which involves a host of researchers. In addition to Wunsch and Marshall, the group using NPACI systems includes associate professor Jochem Marotzke, principal research scientist Detlef Stammer, postdoctoral researcher Ralf Giering, and research engineer Chris Hill.

    "The sparsity of ocean data has been the major stumbling block to progress, but this impediment is being greatly lessened by both satellite data and field data gathered in the international World Ocean Circulation Experiment," Wunsch said. "Moreover, improvements in computation and new approaches to constraining both the models and the observations now make it possible to do for oceanography what has been done for 35 years in meteorological modeling."

    If the state of an OGCM can be brought to full consistency with a variety of global data sets, Wunsch noted, the resulting circulation estimate can be employed to study the consequences of the circulation and its temporal variability on a host of oceanographic problems. The quantitative combination of an OGCM with observations can also be viewed as an initialization of the model--an essential step in climate forecasting.

    "Simultaneously we can make quantitatively useful estimates of the uncertainty of the results and their sensitivity to observational strategies," Wunsch said. "These are important elements in determining what is known about climate change and of utmost importance in designing future observational programs to reduce the remaining uncertainty."


    Consistency with Ocean MeasurementsFigure 2: Consistency with Oceanographic Measurements

    Changes required in the estimates provided by the US National Center for Environmental Prediction (NCEP) in the 10-day averaged fresh water flux field (top) and the heat flux (bottom) as they emerge from the optimization for TOPEX/POSEIDON repeat cycle 21 (early September 1993). These changes are acceptable within plausible guessed errors for the NCEP product and are required for consistency with the oceanographic measurements used to produce the results in Figure 1.

    FORWARD AND ADJOINT MODELS

    "The juxtaposition of model and observations also helps to drive model improvements, such as the representation of convection, eddies, and mixing," Marshall said. These and other ocean phenomena are of intense interest to his group.

    The MIT group works with two models: a forward model and its adjoint model. Both are required for the estimation procedure. The forward component is the Marshall OGCM, which is specifically coded for optimal use of modern computer architectures.

    For the present purpose of a "proof of concept" a global model with 2ö horizontal resolution and 20 levels in the vertical (running from 12.5 to 4950 meters in depth) was used. The global model extending from 80ö in latitude, was started from January potential temperature and salinity fields, after a dynamic adjustment of approximately 1 month. Forcing fields were drawn from the 1993 twice-daily National Center for Environmental Prediction (NCEP) analyses after averaging over 10-day periods. The model has an implicit free sea surface and convectively adjusts vertical instabilities in the density field, according to Stammer.

    An adjoint model is based upon the forward model's code, but describes a space "dual" to the physical ocean. Adjoint models are powerful tools for studies that require an estimate of sensitivity of model output (such as a forecast) with respect to input. The forward model by itself predicts the observations. Misfits between model and observations represent new information in the latter, and the adjoint of the original model is employed to describe the gradients of an objective function with respect to the forward model parameters and control variables.

    Generally the coding of a complex numerical model's adjoint is extremely time consuming and difficult, comparable in effort to developing the forward code. However, the modern computer code of the MIT model, constructed by Hill, made it possible to obtain the adjoint component from the forward code semi-automatically by using the Tangent Linear and Adjoint Model Compiler (TAMC) which was written by Giering.

    "In practice, this system of automatic adjoint code generation has proven to be extremely flexible," Giering said. "It permits easy regeneration of the adjoint code whenever a change in the forward code, including the objective function, is made, and it is easy to add observations and related additional constraints during the estimation procedure."

    THE EXPERIMENT

    The OGCM was constrained by sea surface height data from the altimeters aboard the TOPEX/POSEIDON (T/P) satellite, as well as by surface forcing field (wind stress, heat, and freshwater fluxes) data from the National Center for Environmental Prediction. After iteration with the adjoint model, the results differed from both the forward model and the original data, and they were more consistent with the annual mean climatology (in terms of potential temperature and salinity). These initial results suggest that the adjoint procedure is sufficiently robust to warrant further studies with a broader array of input data.

    "These were computationally intensive calculations on the CRAY T90 at SDSC," Wunsch said. "Our ultimate objective is to be able to conduct such calculations on SMP clusters, reserving still larger calculations for the teraflops machines of the future."

    The forward model development was made possible in large part by a collaboration of Marshall with Arvind of the MIT Laboratory for Computer Sciences. This latter group is also involved in a collaboration with Digital, Sun, Compaq, and Intel to explore the SMP alternatives. They also hope to continue their efforts within NPACI, according to Wunsch. They recently became part of an Earth Systems Science thrust area project on ocean circulation and climate with collaborations in both the Data-Intensive Computing and Programming Tools and Environments thrusts.

    "This initial computation showed that the problem of rendering a GCM consistent with a global scale, time varying oceanic data set has actually been solved--a major step forward in oceanography," Wunsch said. "We are now moving a rapidly as possible to obtain more realistic model resolution, and to incorporate a much more complete set of oceanographic observations spanning nearly a decade of the evolving ocean." --MM