the world have their own models with
their own strengths and weaknesses.
The National Oceanic and Atmospheric
Administration (NOAA) part of the US
Department of Commerce based at Silver
Spring, Maryland, is also in the process of
upgrading its computer systems. Two new
12-petaflop Crays will be moved into a
facility in Virginia and operational by 20 22.
They join existing Cray and Dell systems to
give the agency a capacity of 40 petaflops
to direct into forecasting and research.
This, combined with the Met Office and
others worldwide, means huge amounts
of processing power are being employed
to tell us what the weather will do next. So
we must be getting pretty accurate?
“As your computer becomes more
powerful, your accuracy increases,” says
Roberts. “We tend to see an increase of
about a day per decade in terms of forecast
accuracy. It’s tricky because it depends on
what parameters you’re looking at. One
that we use quite often is the next day’s
temperature, and that’s in the region of 95
percent accuracy for knowing tomorrow’s
maximum temperature.”
The Met Office and Exeter University
have partnered with Google to build an
AI model capable of predicting whether
it will rain up to 90 minutes beforehand.
Google’s involvement is through its
DeepMind offshoot, the people whose
AI trounced all-comers at a StarCraft
II tournament in 2019. DeepMind’s CEO
and co-founder is Demis Hassabis, the
designer of the Theme Park video game, a
UK government adviser, PhD in cognitive
neuroscience, pentamind champion at the
Mind Sports Olympiad four years running,
and yet still only 45 years old.AI ASSISTANCE
DeepMind and the Met Office trained
a generative adversarial network
(GAN) to predict rain across a region
measuring 1 ,53 6 x 1 ,280km ( 954 x 795
miles). To do this, it was shown three
years’ worth of radar data gathered at
five-minute intervals and was asked to
predict what would happen next. A panel
of 50 meteorologists then ranked its
predictions compared to those produced
by more conventional means. The GAN
was ranked first in terms of accuracy and
usefulness 98 percent of the time.
“Rain showers and thunderstorms are
tricky to predict exactly,” says Roberts.
“What our models do well is give a
forecast of areas that are more likely or
less likely to see them. A good analogy is
boiling water. You know it’s going to boil,
but saying exactly where the bubbles are
going to surface is difficult, and that’s
where artificial intelligence may come in.”The project remains an interesting
piece of research for now, however, and
isn’t being put into frontline forecasting.EARTHQUAKES & SEISMOLOGY
Sometimes, however, a forecast becomes
a real-world life-saver, as in the case of
the earthquake early warning system
spread across California, Oregon, and
Washington, with Utah expected to be next
to join the system. Known as Shakealert,
it can identify and characterize an
earthquake a few seconds after it begins—
no mean feat in an area where the ground
is constantly shifting.
Once an earthquake is identified and
the system put into action, it relies on the
laws of physics to alert people before the
quake reaches them: as mobile phone
signals fly at around the speed of light,
which is much faster than the earthquake
waves propagate through the Earth,
people far enough away from the quake
get enough warning to take cover, or shut
down industrial equipment. Trains canstop, airplanes abort a landing if safe, or
stop taxiing. Cars have time to get away
from tunnels or bridges.
A large earthquake that occurs at the
point fault lines running north-south in
California take a quick diversion west and
head out into the Pacific near the home
of the Humboldt Redwoods in northern
California, is felt in San Francisco a
minute later, and Seattle four minutes
after that. It’s not a great deal of warning,
but it can save lives. Unfortunately, there
is always a region close to the quake
where the cellphone alerts don’t reach
people quickly enough.
That said, earthquakes rarely come
along, and while it’s not possible to
predict where a new one might go off
until it’s actually started, the aftershocks
that follow it can be. “The bigger the
earthquake, the more aftershocks it has,
it’s a well-known relationship,” says Dr
Morgan Page, a statistical seismologist
at the US Geological Survey based in
Pasadena, California.An aerial view of the Carrizo plain, South California, with the San Andreas fault easily visible.predicting the unpredictable