The Economist January 15th 2022 67
Science & technologySeismologyAnd now, stay tuned for the
earthquake forecast
O
ne of thequestions most frequently
asked of the United States Geological
Survey is whether earthquakes can be pre
dicted. Their answer is an unconditional
“no”. The relevant page on the agency’s
website states that no scientist has ever
predicted a big quake, nor do they know
how such a prediction might be made.
But that may soon cease to be true.
Though, after decades of failed attempts
and unsubstantiated claims about earth
quake prediction, a certain scepticism is
warranted—and Paul Johnson, a geophysi
cist at Los Alamos National Laboratory, is
indeed playing down the predictive poten
tial of what he is up to—it is nevertheless
the case that, as part of investigations in
tended to understand the science of earth
quakes better, he and his team have devel
oped a tool which might make forecasting
earthquakes possible.
As do so many scientific investigations
these days, their approach relies on artifi
cial intelligence in the form of machine
learning. This, in turn, uses computer pro
grams called neural networks that arebased on a simplified model of the way in
which nervous systems are thought to
learn things. Machine learning has
boomed in recent years, scoring successes
in fields ranging from turning speech into
text to detecting cancer from computer
isedtomography scans. Now, it is being
applied to seismology.Slip-sliding away
The difficulty of doing this is that neural
networks need vast amounts of training
data to teach them what to look for—and
this is something that earthquakes do not
provide. With rare exceptions, big earth
quakes are caused by the movement of
geological faults at or near the boundariesbetween Earth’s tectonic plates. That tells
you where to look for your data. But the
earthquake cycle on most faults involves a
process called stickslip, which takes de
cades. First, there is little movement on a
fault as strain builds up, and there are
therefore few data points to feed into a
machinelearning program. Then there is a
sudden, catastrophic slippage to release
the accumulated strain. That certainly
creates plenty of data, but nothing particu
larly useful for the purposes of prediction.
Dr Johnson thus reckons you need
about ten cycles’ worth of earthquake data
to train a system. And, seismology being a
young science, that is nowhere near possi
ble. The San Andreas fault in California
(pictured), for example, generates a big
earthquake every 40 years or so. But only
about 20 years (in other words, half a cycle)
of data sufficiently detailed to be useful are
available at the moment.
In 2017, however, Dr Johnson’s team ap
plied machine learning to a different type
of seismic activity. Slowslip events, some
times called silent earthquakes, are also
caused by the movement of plates. The dif
ference is that, while an earthquake is usu
ally over in a matter of seconds, a slowslip
event can take hours, days or even months.
From a machinelearning point of view
this is much better, for such an elongated
process generates plenty of data points on
which to train the neural network.
Dr Johnson’s classroom is the Cascadia
subduction zone, a tectonic feature thatAn intriguing new approach to predicting seismic events shows promise→Alsointhissection
68 ClimatechangeandtheArctic
69 Xenotransplantation
69 Theworld’sbiggestfishhatchery
70 Omicron and immunity