Weather forecasting: Expect more and better quality predictions, thanks to supercomputing

Sunday, October 5, 2014

Supercomputing resources have grown at a phenomenal rate in recent years, allowing for much more precision in the modelling of the complex systems which define our weather

by PHIL CHADWICK

In 1976, as a young meteorologist, I was told to "find something else you would like to do besides forecasting as this will all be done by the computer within 10 years." I thought that this was an unfair and adversarial statement. I regarded the computer as a useful tool to do better work – and I still do.
Time passed. I spent 37 years providing forecasts and training. My specialty was severe weather, and the human skill sets of pattern recognition and problem solving were vital to analyze and diagnose those problems. Safety and security was the mission and the human was required to provide those products.

Numerical Weather Prediction (NWP) was also in its infancy in the 1970s. NWP is a collection of mathematical models of the atmosphere that use the current weather and other environmental initial conditions to predict the future weather numerically. At the start, a keen meteorologist could easily outperform the NWP if one chooses to be competitive instead of co-operative about it. Much has changed.

The primitive equations based on physics are still at the core of all NWP, but they are not so primitive anymore. Thanks to increases in computing power over the years, the approximations that are necessary to solve these weather prediction equations have been greatly reduced. Huge advances in satellite technology have much improved the determination of initial conditions.   

Miniaturization of the computer chip has greatly increased computer power. The density of transistors and integrated circuits on the chip has doubled every couple of years– until recently. This was Moore's Law named after Gordon E. Moore, co-founder of Intel Corporation. Moore described this trend in 1965 and it has proven to be accurate, partly because those are the Intel targets they aim their research and development toward.

However, miniaturization has maxed out and Moore's Law can't be sustained for a variety of reasons. Currently, computer chips are "water cooled" by bringing water right next to the chip; air cooling is no longer sufficient to keep the temperature down to allow chip operations. A chip that "loses" its cooling melts within seconds in modern large supercomputers.

Meteorological services around the globe all use supercomputers now. The performance of meteorological services is closely tied to their computing power. There are several performance indicators used to assess this supercomputing capacity. The most common and historical quantity is a value called Rmax, which is a measure of the raw computing performance when executing specific, standardized software (called benchmarks). The Rmax values shown in Graph A are provided in Gflops, or billions of floating point operations/second. With the advent of faster and more powerful supercomputers, they are now often reported in Tflops, or 1012 floating point operations/second; or even in Pflops, or 1015 floating point operations/second. For continuity and ease of comparison, the Rmax values in the graph below are all represented in Gflops.

Supercomputing resources have grown at a phenomenal rate since 1993. For Environment Canada, the supercomputer power has grown from an Rmax value of 20 Gflops in 1993 to 425,000 Gflops in 2014 – a factor slightly over 21,000. The 2011 IBM upgrade installed for EC is a remarkable jump in supercomputing capacity compared to the computing capacity of the previous system.

NWP has made full and prudent use of this increased computing power. Early models had to approximate many of the complexities of the ocean-earth-atmosphere system in order to produce a product in a timely fashion. Those approximations have been gradually and largely reduced over the years. Increased resolution in time and space has allowed for much more precision in the modelling of the complex systems which define our weather. The advent of numerous and much better weather and environmental satellites that constantly probe the atmosphere have allowed a much improved determination of initial atmospheric and other environmental conditions, most especially over the oceans.

Graph B shows the trend in NWP Forecast Quality since its inception in 1958. The Quality Score (110-S1) along the left axis is sensitive to the forecast placement of the high-and low-pressure areas and the associated pressure gradients between them. The Quality Score does not explicitly measure the accuracy of the placement of the associated weather. That is a much more technical verification process but, for simplicity, the placement of forecast pressure patterns is very closely related to the weather.

A Quality Score of 100 would be a perfect match of the pressure centres and their associated gradients between what was predicted and what was later observed. The different coloured lines are for different prediction lead times. As would be expected, the Quality Scores are lower for predictions further into the future. Graph B covers the prediction periods with up to five days (120 hours) of lead time. Graph C depicts the same information, but for predictions out to 10 days.

The labels at the top of each column are the names of the supercomputer systems used during the respective time periods. The CYBER 76 was the tool I used (sparingly) when I joined the Atmospheric Environment Service in 1976. The IBM P5 provided the information I consulted extensively when I retired in 2011.

NWP is a very complicated process, but the trends are clear. The Quality Scores of the predictions are generally on the upswing. The large improvements achieved in the Forecast Quality between 2001 and 2003 are largely due both to better data assimilation techniques and the injection of satellite data from new sensors. More satellites are scheduled to be launched and our analysis of the initial conditions of the earth-atmosphere system is destined to continue to improve.

Currently the four-day (96 hour) forecast is as skilful as the three-day (72 hour) forecast produced a decade ago. The trend has been that NWP adds about 24 hours of predictability every decade.

The question remains how long this trend can continue until the "butterfly effect" swamps improvements in meteorology. In chaos theory, the butterfly is the unknown, small and seemingly insignificant flutters in the initial conditions of the atmosphere. These unmeasured flutters do not register into the initial conditions and cause the patterns to evolve along a different path of weather from that predicted by the NWP. Ensemble forecasts (see "Ensemble forecasting – a better tool for today's farmers," Better Farming, February 2014) allow one to gauge the sensitivity of the evolution of the current atmosphere to those small unknowns. All is not lost if we use these ensembles and treat the weather as a range of probabilities versus a single, deterministic certainty. Probability of Precipitation (see "How best to interpret those precipitation forecasts," Better Farming, April 2014) is a familiar example of the use of probability in forecasting.

Training and education are the foundations required to provide any quality service. During my initial meteorological training, I was instructed on how to use Venn Diagrams (see Figure A) to aid in my forecasts – yes, Venn diagrams! Satellite data was just coming online in the late 1970s and I quickly realized that weather did not arrive in the intersections of overlapping circles. The actual weather was much more complicated than that. The new satellite patterns were so complicated that at first they seemed undecipherable.

The satellite became my tool of choice to augment the NWP data of the day. I developed powerful techniques to interpret the lines and swirls revealed in the satellite imagery. This imagery was like a book describing the actual state of the current earth-atmosphere system. One just had to develop and learn the vocabulary.

The conceptual models based on those atmospheric signatures allowed me to interact positively with the NWP guidance. Comparisons between the satellite atmosphere and the NWP revealed when and where the NWP was missing something. The human could then step in to make the required fixes and everyone was happy.

The NWP continued to improve dramatically during my career, but there was still plenty of room for improvement using my human-friendly conceptual models. This was especially true during the big events when the atmospheric systems diverged dramatically from the climate normals. Those were the forecasts that could really contribute to our mission of safety and security – when the weather you experienced was wildly different from the weather you expected or the climate. Does anyone remember the Ice Storm of 1998?

The research completed over the years has been designed to empower meteorologists to fully use their human pattern recognition and problem-solving skills. Much of this material has been published under the brand of "The Satellite Palette." I regarded these approaches to the weather in the same way that I approached my canvases when I painted. Science may be regarded as an art and vice versa. These conceptual models and a wealth of other environmental training can be found for free through the Cooperative Program for Operational Meteorology, Education and Training (COMET) and the following website: https://www.meted.ucar.edu/.

So what can we expect in the next 40 years? The meteorologist-machine blend remains the future of weather forecasting. The goals of both are to provide superior predictions upon which clients will base their decisions and take appropriate actions. Neither the meteorologist nor the machines are perfect, but together they can produce something of real value.

The very best is achieved when the strengths of the human and the machine are creatively combined. There is still room for a well-trained meteorologist to add significant value to the forecast, contrary to the challenging statement made to me as a young and keen meteorologist. The role of the forecaster for the next 40 years will also evolve to provide more specialized services, consultation and services not yet imagined. Meteorologists will help clients make the best use of weather information for their needs. Supercomputers will never talk to clients and users. Forecasters will.

The requirements of modern society for accurate and timely messages of significant weather should be expanding as fast as our capabilities to provide that highly detailed information. Improved methods to communicate all of the information are certainly required. Weather icons, although attractive, do not convey much if any information – sample, http://thumbs.dreamstime.com/z/weather-icons-9504363.jpg.

Weather-related insurance costs in Canada have exceeded $1 billion since 2007 and this figure will continue to soar. The wise use of weather information and taking action to mitigate weather and climate threats is the best and perhaps only way to bring those costs down.

Everyone may talk about the weather, but meteorologists, highly specialized researchers and computer scientists are actually doing something about it. My good friend Jean-Guy Desmarais, also retired from Environment Canada, shares this vision and provided some of the important information and graphs included in this article. The future of weather forecasting is very exciting. BF

Phil the forecaster Chadwick has been a professional meteorologist since 1977 specializing in training, severe weather and remote satellite and radar sensing.

Four decades of dramatic change
Over the last 40 years, the changes in NWP have been much more dramatic than the evolution of humans in the same time period. They include:

• Large increases in supercomputing capacity allowing much more sophisticated data assimilation techniques to achieve a better representation of the initial environmental conditions.
• More sophisticated prediction models using more accurate depictions of the physical processes involved both in the atmosphere and in the layers where atmosphere interacts with the underlying surface. Fewer approximations are also paramount to better simulation of the real world processes.
• And probably the most important, the launch and operational implementation of a large array of weather and environmental satellites over the last couple of decades. Much better and more diversified sensors aboard these satellites now probe the atmosphere constantly, and several millions of these satellite observations now find their way into modern NWP systems.

The result is a much more accurate depiction of initial conditions required to make predictions, most especially over oceans. Every day, several billions of satellite data information pieces make their way to major weather centres, where they are databased and processed. Every six hours, several tens of millions of atmospheric profiles for several parameters are generated from these datasets and ingested in data assimilation systems, hence leading to much better representations of the initial conditions of the environment.

This is a huge change compared to the old days (until the mid-1980s), when a few tens of weather ships scattered over the oceans were the only data sources for the provision of 3D initial conditions over the oceanic areas. It is not too surprising that NWP forecasts in those days were suffering from major flaws when even the atmospheric initial conditions were ill-defined. Considering that 70 per cent of the earth's surface is covered by oceans, satellites now provide valuable data over important regions where data was largely missing. BF