without knowing the distribution strength
deep underground, or to image it, we’re
not going to be able to predict how big it
will be until it gets to that size.”
And it’s not just in Pasadena where
the idea of turning machine learning
loose on earthquake data has taken root.
In 2018 , a competition was launched in
which entrants had to design their own AI
to predict earthquakes being artificially
generated in a laboratory. It was open to
anyone, and thousands of people entered.
The results, published in February this
year in the Proceedings of the National
Academy of Sciences [https://bit.ly/MPC-
wobble], were remarkable. Some teams,
none of them with any background in
seismology, nearly got it right.
It’s important to remember that this
was an artificial situation, a synthetic
benchmark if you like, but the competition
not only yielded insights into seismic data
but is also being used as a tool to further
spread the good news about machine
learning [ML] in geophysics. “The
approach may provide a model for other
competitions in geosciences or other
domains of study to help engage the ML
community on problems of significance,”
the organizers, a coalition of scientists
from Los Alamos National Laboratory,
Stanford University, and Google, wrote in
the abstract to their PNAS paper.
There are definite signs the lab
data can be applied to the real world.
“ML has revealed the time remaining
before an earthquake in the laboratory
and particular types of tectonic
earthquakes known as slow slip events
can be anticipated from statistical
characteristics extracted from seismic
data,” the organizers continued.
At first, nascent AIs were provided
with huge amounts of data about the
movements of an artificial fault—two
steel blocks surrounding a lump of
rock that are moved to simulate the
strains of being squashed between two
tectonic plates. If the rock cracked, then
the whole assembly lurched, and that
was considered an earthquake. Pretty
soon, the machine learning models had
identified a handful of energy signatures
in the data that always preceded a crack.
This was considered too easy, so the
team moved on to Kaggle, a platform for
sharing ML research, and opened up their
competition. Putting up a $50,000 prize
for the top five teams, they announced
they would be simulating earthquakes for
six months in 2019 , and invited all comers
to predict them. To the ML community,
this was too good an opportunity to
overlook. Over 4 ,500 teams signed up,
and the eventual top five became really
good at quake prediction, despite some
unorthodox approaches. One fed random
noise back into the prediction model,
which made it better. No one knows why.
The winning team, The Zoo, was
made up of eight people across the US
and Europe who had never met, and
who had expertise in AI, mathematics,
computer science, operations research,
electrical engineering, insurance, hotel
management, and credit risk. Anything
but seismology. The Zoo never took first
place in any of the competition’s daily
rankings but ranked consistently high
enough to take the overall win. Hopes are
high that the lab data will scale up to San
Andreas level, but first, it needs to be tried
out on relatively quiet faults to see if the
ML models hold up out of the laboratory.
And it’s not just earthquakes that
this technology could learn to forecast.
“There are a lot of earthquakes in volcanic
regions,” says Page. “And when volcanoes
pick up, a lot of earthquakes start too.”
So in the future, we could be looking at a
net of AIs monitoring the globe, looking
out for severe weather, quakes, volcanic
eruptions, and whatever else can be
modeled from the enormous streams of
data our sensors pour out day after day.”
The science of prediction has come a
long way from the invention of the weather
forecast in the 186 0s, but we probably
could have predicted that it would.
The 1989 San Francisco earthquake caused the Cypress Viaduct to collapse. Earthquakes are much more difficult to predict than the weather.
HOL 2021 MAXIMUMPC 43
©^
US
GS
/P
UB
LIC
D
OM
AIN
,^
