Page 18 - EngineerIT May 2022

ICT SUPERCOMPUTING

Usual data sources could include satellites, weather stations and balloons, delivering
anything between 500 gigabytes and one terabyte. But before the data can be used, it
must be put through a process of quality control. Once that process has been completed,
mathematical models are then used to make forecasts. Known since the 19th century,
these are equations that describe the state, motion and time evolution of various
atmospheric parameters such as wind and temperature.
The good news for the continent is that scientists are working on addressing our
data issue.
In Narok, Kenya, for example, climate scientists, local meteorologists and farmers
are working on a potential solution that could be critical to increasing data collection
on the continent.
The idea is to install simple, relatively inexpensive weather stations like the one at
Ole Tipis, a boarding school just outside of the town of Narok, across the continent. The
Ole Tipis station is one of 115 in Kenya run by the Trans-African Hydro-Meteorological
Observatory (TAHMO), which has a network of 626 stations in 20 countries. If they
get this right, it could positively impact Africa’s participation in understanding climate
challenges as well as its ability to prepare for them. Jim Holland

Massive compute power Set Off a Tornado in Texas?’
Turning data equations into accurate forecasts requires an additional factor – compute Setting aside chaos theory, there’s
power. To understand how this works in practice, it makes sense to use a simple another reason why forecasting may
illustration. If the United States were divided into a mesh of 10km blocks, then a certain take time to become more accurate –
level of compute power would be needed to provide localised forecasts inside each block. the science itself. Although compute
The difficulty arises, however, when the size of the blocks is reduced. Thunderstorms, power doubles every two years or
tornados and smaller scale effects are very much linked to local weather, and with a so, weather science takes longer to
large mesh it’s easy to miss them. It’s similar to being a fisherman – a much denser net is catch up. Supercomputers were first
needed to catch small fish. used in the US in the 1960s and 70s,
but it took between ten and 20 years
Enter AI for forecasts to become much more
Because of the enormous amounts of compute power involved in making these accurate.
calculations, scientists are now looking at how other technologies like artificial Still, the compute power that is
intelligence can improve forecasting. now available has massively improved
Instead of using brute-force computation to forecast weather based on present forecasting. When weather predictions
conditions, AI systems review data from the past and develop their own understanding were first made in the 1950s, the results
of how weather conditions evolve. And they are already having a significant impact on were highly inaccurate because of the
forecasting. For example, the UK’s Meteorological Office recently carried out a pilot of limited computational power available.
AI technology to predict flash floods and storms. Using radar maps from 2016 to 2018, To give an example of how things have
the system was able to accurately predict patterns of rainfall in 2019 for 89% of cases. moved on, a weather model that would
Advancements in technology mean its four-day forecast is now as accurate as its one-day have taken 600 years to run on computer
forecast was 30 years ago. systems in the 1960s now takes just
15 minutes on a standard Lenovo
Bumping up against the Butterfly Effect ThinkSystem server.
New technologies will undoubtedly usher in an era of more accurate forecasting, but There is every reason to believe
they will never be able to make long-term predictions about the weather with 100% that as compute power increases in the
accuracy. That is because the equations that are used to make weather forecasts are next few years alongside our scientific
non-linear – they have a degree of chaos embedded in them. knowledge of weather patterns, it
As early as the 1960s, Edward Lorenz, an MIT meteorologist, was arguing that will be possible to make even more
it was fundamentally impossible to predict the weather beyond ten days. Central accurate predictions. And with the
to his argument – which later became known as chaos theory – was the claim that ability to predict extreme weather,
small differences in a dynamic system like the atmosphere could trigger completely supercomputers have the power to save
unpredictable results. The most famous formulation of this theory was Lorenz’s lives and make a profound impact on the
1972 academic paper ‘Predictability: Does the Flap of a Butterfly’s Wings in Brazil world. n

EngineerIT | May 2022 | 16