past, and running it forward. We have a
model that runs for all in California and
has all the faults in it. We tried to run the
physics on the fault, but because we don’t
actually know the state of stress deep in
the Earth, we’re not running the current
state forward in time, but running other
versions of what history could be and then
mining those catalogs for data. We don’t
yet have a good handle on how similar our
six base models are to the actual Earth,
they are more like guesses. Weather
people can measure a lot more in the
atmosphere than we can underground.”
Indeed, while the Met Office can rely on
a network of weather reporting stations on
land and sea across the world, all within
the atmosphere they’re trying to study,
the USGS has to make do with boreholes
and surface seismometers—not ideal if
you’re looking for things happening an
average of 12 miles below the surface in
California, and over 350 miles down in
parts of the Pacific near Japan.
“We’re only seeing what’s happening
on the surface, boreholes aren’t that
deep,” says Page. “We measure surface
movement using GPS, and how quickly
the Earth is moving, but it’s just the
surface extrapolating what’s going on
underground. We can figure out when and
where an earthquake happens because
we know the travel time of the waves.
Similarly, when a big earthquake happens,
we try to image it underground, based on
our measurements on the surface. It’s like
having an MRI scan on your head with the
sensors on the outside, and they manage
to image what’s going on in your brain.
“The model is only as good as our
understanding of the Earth’s structure,
and how fast the waves are propagating
underground. There are always
uncertainties. The epidemiological-type
models came from the epidemiological
community, and in the 198 0s were ported
over to seismology. They’re still the best
models we have for predicting the rates
of earthquakes, and we’re still plugging
away to incorporate more physics into
these models. We’re seeing hints of what
lies beyond those simple models, another
layer of predictability waiting to be seen.”
Luckily, different places around the
world can be studied to provide further
insight, as some are more enthusiastic
than others in terms of the number of
aftershocks they produce. “There are
different tectonic regions,” says Page,
“so there are places over continental
crust, and there are places over oceanic
crust, and there are subduction zones or
one place that’s affecting another, and all
these places have different aftershocks.
Some have depleted aftershocks andothers have more energetic aftershock
sequences. We try to get as much data as
we can by finding earthquake sequences
in similar regions so that we can get the
best estimate of what’s going to happen.
“Subduction zones [where cold ocean
floor sinks back into the hot mantle of the
Earth] are the most energetic, and the
least energetic are transform boundaries,
like the one in the middle of the Atlantic.”USGS & MACHINE LEARNING
The USGS also has some pretty interesting
hardware. The first supercomputer
available to all USGS staff was Yeti, a Cray
with 1 ,4 56 Ivy Bridge cores and 2 ,272
Haswell cores, 5, 120 Nvidia Quadro K2 200
GPU cores, and 2 9, 952 Tesla K 80 GPU
cores, plus 366 Xeon Phi cores. Yeti has
32, 12 8GB of RAM and was benchmarked
at 1 ,02 1 teraflops in 2019. Elsewhere,
you’ll find Tallgrass, a Cray Urika-CS
prototype designed for AI and analytics.
It features 792 CPU cores, 122 ,880 CUDA
cores, and 15 ,360 Tensor cores for a total
capacity of 236 teraflops. There’s Denali,
another Cray with 9 ,2 80 CPU cores
and an estimated peak computational
performance of 660 teraflops when it
becomes operational in 2022.
Machine learning is starting to be used
to tackle the huge amount of data the
USGS seismometers produce. “It mines
more earthquakes out of the catalogs
than a human analyst could find,” says
Page. “A lot of it is template-matching,
so it figures out there’s one really small
earthquake, a bunch of wiggles basically
on the seismogram, and then it combs
through years and years of continuous
data looking for other little wiggles that
more or less match that. And from that, it
can find even smaller shapes that look like
those wiggles. The better catalog we get,
the more patterns we’ll be able to figure
out. If we can go from magnitude two to
magnitude zero that’s 100 times as many
earthquakes, which is a big difference.”
“From the data we have so far, it
seems like big earthquakes and little
earthquakes start off the same, soEARTHQUAKE
DEEPDIVE
In geologically active areas, the
ground can move a lot without it
becoming an earthquake. The trick
of earthquake detection is to tell the
difference between the earth shifting
around a faultline, including tremors,
and the destructive shaking that
accompanies a quake.
During an earthquake, the
rupturing fault sends out three
different types of waves. P-waves
come first, setting off earthquake
detectors, but rarely causing
significant damage. These travel
faster than other seismic waves
and can travel through solids and
liquids—even gasses, but they
are rarer underground. The P can
stand for ‘primary’ or ‘pressure’
wave, as it’s created by repeated
compressions and decompressions.
Their motion is in the same direction
as the wave propagation.
S-waves come next, and these
are the waves that cause the most
damage. The S can mean ‘secondary’
or ‘shear’, and it’s the shearing or
side-to-side movement (their motion
is perpendicular to wave propagation)
that will knock down buildings.
Raleigh waves, triggered by the
interactions of P and S waves at the
surface, are slower still, but have
a rolling motion similar to ocean
waves, and can cause even more
damage. Luckily, their amplitude
decreases exponentially with the
depth of the earthquake’s hypocenter
underground, so only strong,
shallow quakes produce particularly
destructive quantities.
Raleigh waves operate at a
frequency that, though inaudible,
can be detected by mammals, birds,
insects, and spiders. This is an area of
active study, but it’s thought animals
may receive an early warning of
quakes through this system, which is
used by elephants to communicate at
great distances through infrasound,
though this operates at a much higher
frequency than the waves generated
by earthquakes. Humans have nerve
receptors in their joints, Pacinian
corpuscles, that should be able to
detect Raleigh waves, but people do
not seem to consciously respond to
the signals. It is not yet known why.A kinemetric seismograph responds to
movement by tracing on a roll of paper.predicting the unpredictable