ECMWF hosts Workshop on Teracomputing for meteorology

On average there were 16 presentations a day crammed with the technical aspects of how to partition data across large parallel supercomputers and then manipulate it while doing computation using mathematical models which describe weather patterns over time. As a saving grace these were interspersed with weather maps and video pictures from simulations illustrating for example, the recent flooding in continental Europe and the UK at the end of October and beginning of November. Computer simulated results obtained 7 days and even finer ones obtained 2 days in advance confirmed the weather patterns from satellite pictures taken on the days it rained.

The flooding caused enormous economic damage and hardship to many people. But for meteorologists, this stormy cloud had a silver lining as it confirmed yet again the accuracy of their weather prediction models. The recent storms over Europe are expected to increase in frequency and intensity as global warming begins to bite and this is likely to make life even more exciting for weather forecasting buffs as the prowess of their models is likely to be tested to destruction. Thus short and medium-range weather is of great interest to us all, as it has enormous economic influences on our every day lives.

Resources needed for weather prediction

Present-day weather forecasts are made almost entirely by using computer modelling techniques. This requires a sophisticated high technology infrastructure and enormous resources for measuring and collecting weather data from across the world. The advent of parallel vector supercomputers with large memories and massively parallel distributed and shared memory systems has made it possible to forecast the weather of the whole world several days ahead with reasonable accuracy.

Statistics show, that today, the 7-day forecasts from ECMWF, in the UK, at 60% accuracy are better than the 3-day forecasts made in 1972, at the start of reliable numerical weather forecasts based on modelling with computers. Global forecasts of medium-range weather patterns are issued for 3 up to 10 days ahead.

In addition, supercomputers are used to predict important weather elements directly. Rain and snow precipitation predictions are made routinely up to 2 days ahead using a regional model with a higher resolution. It predicts the same variables as the global model but it additionally predicts cloud, water and turbulence intensity. This allows much better definition of detailed weather patterns.

The atmosphere is modelled from the earth's surface to 65 kilometres in the vertical, so there are many variables which influence the state of the weather at any given time. Dynamics of sun's radiation and simplified chemistry in the stratosphere, the interactions of radiation with cloud energy and water cycle in the troposphere, ocean surface waves, ocean circulation, simplified sea ice behaviour, land snow and freezing soil moisture, soil moisture and simplified vegetation, are all included in the model.

Medium-range forecasting has a different role; charting long distance sail routes, avoiding weather trouble spots minimising safety risks from stormy weather, saving time and fuel. It can also be used to anticipate potential flooding , snow precipitation suitable for skiing, or predict electricity consumption a week in advance. Indeed there are many other uses which medium-range weather forecasts are profitably utilised; it even includes supermarkets, which use them to decide whether to order extra salad and ice cream when hot weather is predicted.

For medium-range weather the model forecasts the wind, temperature and humidity at over 4 Million points throughout the atmosphere and other weather parameters (snow, soil moisture....) at around 140 thousand points on the earthⳠsurface. To make this possible, data is collected from thousands of ships and points on land as well as from weather balloons and satellites at great expense. Indeed, world-wide several billion pounds is spend each year collecting weather data. Thus, in addition to solving the forecast problem, much computer time is spent on data analysis and output. The observed data, which normally has gaps in it, is merged with a "first guess" from the forecast model. Observations containing gross errors are rejected. A wide range of output products are disseminated to national meteorological centres all over the world. The operational forecast suite takes several hours of computing time, so the total amount of floating point operations performed is typically several Petaflops every day and also handles many TeraBytes of data.

Climate simulation as a predictive tool

Other examples of weather model simulations include the development of the strongest natural climate fluctuation, El Nino Southern Oscillation (ENSO), which was successfully predicted 6-12 months in advance, to the benefit of countries not only in the tropical Pacific, but also in North America and other regions of the world. Another major area of success is the considerable improvement in predicting anthropogenic climatic change using enhanced numerical climate models.

Driven by the sun's energy, the climate subsystems - atmosphere, ocean, biosphere and cryosphere (ice and snow) - act together in a complex fashion. Understanding how sensitively the earth's climate system react to natural and human influences and which regional and global climatic changes are to be anticipated in the future has become a litmus test, a measure of how science can help us to survive the 21st century and beyond. Industry, politicians and society are all awaiting solutions based on well-founded principles to guide them when taking decisions in this area.

Whether we are talking about the ban on chloroflurocarbons, the reduction in greenhouse gases or the calculation of premiums in the insurance industry, climate studies supply the fundamental data. Climate studies are currently shifting from modelling the physical climate system to modelling the "total world system", with chemical models and socio-economic models assuming increasing importance. The fastest parallel supercomputer, currently being developed by NEC for the Japanese Earth Simulator programme is expected to be operational and in production by March 2002, and will no doubt make a significant contribution to understanding the total world climate system. The simulation of the virtual earth is validated by data from earth observation satellites. This means that a number of projects in the study of world climate and the environment can be supported.

Supercomputer power currently used for weather models

>From the presentations the supercomputing power used by the large weather and climate centres is between half and one Teraflop/s peak and about 100 to 400Gflop/s sustained. With the exception of centres in the USA which are discouraged by their government, the favoured supercomputer used in weather modelling by the rest of the world is the Cray T3E and Japanese vector processors from Fujitsu and NEC. This is hardly surprising since the NEC SX-5 has a well balanced architecture and high memory bandwidth which is well suited for weather modelling applications.

In the next two to three years weather models are expected to incorporate several new important factors influencing atmospheric changes which occur on very small scale. These include heating of the soil by the sun and air turbulence near the ground, or higher up air flows over mountains. These can never be represented accurately by the models. Their effects are instead accounted by their influence on the behaviour of larger-scale parameters. Future models are likely to expand the number of grid points to more than 20 million. This will bring the sustained computing power needed in the many Teraflop/s range. Computer vendors are gearing their next announcements for this procurement cycle.

The future is bright, bleak or indeterminable

As for the future, silicon semiconductor devices are expected to reach their physical limit at around 2010, but Dr. A. MacDonald, the director of NOAA/ Forecast Systems Laboratory, in Boulder Colorado, USA, projected that computer technology is likely to continue to deliver MooreⳠLaw performance for the next 50 years. By then Quantum computers could become mainstream, and meteorologists will require 10 exaflop/s (10 to the power of 18) for running their global weather and climate simulation models with a 3km grid size.

This is predicated on the assumption that we are able to harness the dire effects of global warming. Judging from our politicians squabbles at Kyoto and this week in the Hague, Netherlands, and the reluctance of nations particularly the USA to implement even modest greenhouse gas targets, the omens are bleak. The USA with just 4% of world population is emitting 30% of pollutants so a load balancing is necessary. One canⴠhelp feeling that the aggressive consumer culture which helped to win the cold war is likely to be responsible in triggering the sixth mass extinction in the earthⳊhistory, so we lose the peace.