The Economist - USA (2022-01-15)

(Antfer) #1
The Economist January 15th 2022 67
Science & technology

Seismology

And 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 scientific 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  are

based on a simplified model of the way in
which  nervous  systems  are  thought  to
learn  things.  Machine  learning  has
boomed in recent years, scoring successes
in fields ranging from turning speech into
text  to  detecting  cancer  from  computer­
ised­tomography  scans.  Now,  it  is  being
applied to seismology.

Slip-sliding away
The  difficulty  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 boundaries

between  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 sufficiently detailed to be useful are
available at the moment. 
In 2017, however, Dr Johnson’s team ap­
plied machine learning to a different 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  that

An intriguing new approach to predicting seismic events shows promise

→Alsointhissection
68 ClimatechangeandtheArctic
69 Xenotransplantation
69 Theworld’sbiggestfishhatchery
70 Omicron and immunity
Free download pdf