Where We Stand in Earthquake Prediction | Marine A. Denolle || Radcliffe Institute

Where We Stand in Earthquake Prediction | Marine A. Denolle || Radcliffe Institute


me[MUSIC PLAYING] – Thank you very much
for joining us tonight. I really want to thank Meredith
for always doing and making this really nice,
elaborate introduction that are at the same time humbling
but extremely flattering. So I really, really appreciate. So tonight, I’m going to speak
about a topic that’s actually quite new to me, but I
thought it was a lot more relevant to humanity. Population is how
we can or we may or we may not be able to
actually predict earthquakes. And if you fly to the San
Andreas or San Francisco and you have your
laptop open and you show seismogram
everybody will ask you, can you really
predict earthquake? When is the next big one? And every time we tend to, we
have to say carefully that we probably cannot
predict earthquake, because we can have liability
issues in saying wrong things. But I’m actually
not losing hope. Recently, we’ve harnessed
so much data and we’re making more observations
that I’m actually– I think that one day
we’ll be able to do so, but I wanted to go
with you and show you what are the
timescales involved in and where it actually matters. And so, this is a
quite old photo. It’s taken after
the 1906 earthquake. It’s a fault that runs through
the city of San Francisco. So that image that you
see here, of course, it’s older constructions,
it was very dramatic. Most of it was due
to the fire that started after the earthquake. So earthquakes do generate
a lot more hazards that are just coming out after that. The other type of– When we think about
earthquakes, we think about this previous photo. And what we see here
it’s mostly damage that is sudden right
after the earthquake. It’s people dying,
it’s building collapse, it’s highway collapse
as you mentioned. And these are iconic figures
of what earthquakes are. What I also want
to point out it’s not necessarily the
casualties that we refer to, but it’s also the
great economic loss that are a component
with these earthquakes. Some countries have really
a lot of trouble recovering from those, such as Haiti in
2010, Japan in 1994 and in 2011 was hit twice badly. So there’s a lot of economic
loss due to these earthquakes. Earthquakes are also
the way that the earth has to form landscapes. This is a photo from 2010,
another big earthquake in Southern California. And you can see
here very nice fault scar that’s not been eroded
yet and a geologist for scale. And if you can see
the hills behind, you actually start
understanding that earthquakes are the cause of the building
of the landscape around us. And I want to zoom out
and look at the Earth. And I can tell you
most of the topography that we see here is
due to earthquakes. So what you can see here is the
[INAUDIBLE] image topography map. We are looking over there,
centering around the Pacific ring, because we have a nice
distribution of seismicity around it. But you can see there’s a
lot of texture in this photo. It is just showing you that
the earth is very dynamic, and actually the earthquakes
are a system, a mechanism to deal with this. So the plate tectonics is
basically all of these. They’re not totally rigid,
but we can simplify the system as a very rigid plate. They are surrounded
by those zones, that you can see the
topography quite clear. You can see the
zones here and here. The topography shows you that
this is the Pacific plate. It’s surrounded, it’s
bordered by active faults, and the San Andreas
fault being one of them. So if you look at the Earth
at this scale you say, these are lines. People like to see fault lines. It’s not lines. It’s the surface, it’s a volume. But we can simplify
them as line, and then we can assume that
maybe earthquakes are on it. Earthquakes are indeed mostly
located on these lines here. So these are ridges, this is
the San Andreas fault system, this is Japan. And you can see
that globally, you may have distributed seismicity,
but all those stars are representing where most of
the big earthquakes are. So crudely speaking, you know
that at boundaries of plates you may have earthquakes. Of course Oklahoma
is showing up here, not quite but the [INAUDIBLE]
that you can have seismicity. But on to the
first order, if you want to know where
earthquakes are, it’s usually on those
plate’s boundaries. Now if you zoom in
to this boundary, it’s actually a
more diffuse system. These red lines
and the black ones are representing maps faults. And you can see
it’s not in line, it’s not necessarily straight. It has curvature, it
has nice delineation. The dots here are
showing the seismicity that we have that actually
seems to correspond to those topography features. So if we want to know
where faults are, traditionally what we look
at is the past seismicity in these catalogs and look
at where seismicity is. So this is San Francisco
and the biggest earthquake was along this fault. And
you can see for most of it, there’s actually no
seismicity at all. So if you rely on these catalogs
to know earthquakes are and try to predict it, you are actually
missing out a huge part, because here you can
have a large earthquake, but you have any seismicity
to understand it. So I spent my PhD, you
know, somewhere over there and had zero earthquake for 5
and 1/2 years, which is good. But sitting there
it was interesting. Faults can be very quiet. So to deal with
this, geologists, they dig big trenches
through faults and they look at the
texture and the composition of the minerals in rocks. And they try to date
those rocks and then to try to find the
history of the faults and how many
earthquakes there were. So this is an example on
the Wasatch fault in Utah, which borders the Salt
Lake City to the east. And so, if you actually
look at those rocks and you make time
history of them, you can find when
earthquakes happen. And I’m showing you the best
example we have of this. So this is New Zealand. And I don’t know
if you can tell, but there’s some
linear feature here. Linear features are not natural. They’re really caused by faults. And this is highlighted here
what we call the Alpine fault. From a French perspective
it was like, these aren’t all the Alps. [LAUGHTER] – [SPEAKING FRENCH] . But New Zealand is like it. They also make excellent cheese,
so they have a good excuse. But so basically,
this is the Alpine fault. It’s a quite fast
fault. It’s also heavily– There’s nice topography to it. And if we trench
through it, geologists have found a nice
series of earthquakes. It’s also nice because
it’s not really populated, so it does not cause
tremendous damage, so we can use the word nice
to think about earthquakes. This is the dates of
those earthquakes, in thousands of years. And you can see this
nice regular line here with [INAUDIBLE]
distribution [INAUDIBLE] uncertainty of
when these earthquakes happen. So if you look at the
interval between each of them, you make those plots here,
which look at the recurrence interval between earthquakes,
and you see that they’re mostly around 300 years. And for all of these 10
to 20 year earthquakes, every 300 years you
may have an earthquake. And so the last one
being in the 1700s, we’re close to the end of the
cycle for this earthquake. We think about earthquake cycles
here because every 300 years, we can have an earthquake. So bingo, we should be able
to predict earthquakes. Of course, I’m not
giving you this talk to say yes, we solved that. Many faults do not
behave like this. This is a time series here
showing the lifelong time series of the faults. These [INAUDIBLE] are showing
you when earthquakes happen. These are several faults. This is the Alpine fault, you
see that beautiful time series of periodic earthquakes. And then you look at
the San Andreas fault and the system is actually
[INAUDIBLE] not regular. It has these weird
clusters of seismicity. The Dead Sea Transform fault has
somewhat of a regular interval, but not at all. And the Dead Sea,
this one is actually showing clusters of events. So the predictability
here is not strong, because we have
clustering of events and we can’t even
control those clusters. So what happened on a fault? We think about
earthquake cycles, because an earthquake
happens, it heals, it’s in between
cycles and hit again and it’s supposed
to be periodic. Scientists today do not
really like the word cycle because of this irregularity
in some of the fault systems. But I’d like to just
emphasize that it’s just a very, if we think about
it, a very simple model. You can think about
this model where you have a spring in a handle. You have a sure phase that could
be rough or smooth as you wish. And then you have a
block, and you pull it. So that if you have elasticity
in your string here, the block is not
going to move as you pull until it’s too
tensed, and then the block will slide it toward you. And so this effect of
increasing the stress loading on the
string, and then slip, and it sticks, and it slips,
and so on and so forth, is the fundamental view
of earthquake cycle, where you increase
the force the plates are moving at a somewhat regular
pace and then it’s locked. It sticks, and then
[INAUDIBLE] to an earthquake and so on and so forth. So if we model
forces in the earth, you can say it’s a constant
forcing, constant slip constant forcing, constant
slip, this is the ideal case. We have in the United
States one fault that seem to behave like this. It’s in San Andreas fault
in the Parkfield area and it was so
regular in the past. These are years of occurrence
of these earthquakes. Every 20 plus years there
was a little magnitude 6 that would occur on this fault.
So these are the earthquakes series and you can look at
number of earthquakes and time. You draw a line and
you predict when the next earthquake will go. So the prediction was that
between 1987 and 1983, we should have an earthquake. So the USGS deployed
a lot of resources around Parkfield to
instrument the fault and to capture the full cycle. But the earthquake
did not happen. And I think a lot
of the seismologists completely lost faith
because it took many more years for the earthquake
to actually occur. And so I think when big events
like this, either like Tohoku or all these events where we
don’t expect them to happen, it’s very hard for
seismologists to react. And so I think because of
that, a lot of the seismology has been going away
from prediction. It was all about prediction
before, this has failed, let’s move on and not
predict earthquake any more. So let’s think about
this time theory that I was just drawing you. We have the earthquake
at time zero, we have what happens before
or in between earthquakes, and then we have
what happens after. So here, I’m showing you
that between years to months to second before
the earthquake, this is the phase that we
call a nucleation phase. We are at the end of
the earthquake cycle, something is happening, we
will have an earthquake. We just don’t know when t=0 is. What happens after is
the earthquake ruptures. It’s not and
[SNAPS FINGERS] instant. It’s not a snap like
this, it takes some time. So we have a few seconds to
tens of seconds, during which and after which all the
shaking, the ground motion goes, and that’s what
destroys the building. So the prediction of
earthquakes within years, I’m a little bit
skeptical today. And I would say that
probably we may not be able to do that
anytime soon, mostly because we do have many
years in our records. But let’s look at nucleation. What happens between months and
seconds before the earthquake? Unfortunately, all of
these precautionary phases are seen or detected
after the earthquake. So far, we have
seen those phases, but it’s always after going
back to looking at the data. So in 2011, March 29, 2011,
magnitude 9.0 in Japan. This is a map of Japan. This plot is showing you with
colors the amount of slip on a fault surface. It’s dipping like this. And the contours are
showing you where the slip during the
Tohoku event happened. So basically in
the middle, that’s where you had about
15 meters of slip and then nothing in between. And these colors
are showing what happened before the earthquake. And so using data
on shore, scientists have been able to look at how
much slip, very slow slip, happened years before
the earthquake. This line here,
this time series, is showing you what a GPS would
record as a function of years. So you can see we
don’t have that many, we have maybe a few
decades worth of data, but not that much more. And if you remove
some of the effects, you can see that there is this
non-constant but accelerating phase of displacement, and
this is mapped into the fault. And so we have this
acceleration of slip, it’s very slow and
steady, but it grows. And we can detect it by
looking at years of data. It’s very often afterwards
that we notice this. If you look at the foreshocks. Foreshocks are earthquakes that
happened before the earthquake, the main shock. Foreshocks are only
called foreshocks after the main shock,
so it’s hard to use them as a prediction. What I’m showing you
here is three earthquakes that happened in
Southern California. This is the El Mayor-Cucapah
earthquake in 2010. I have sensors actually nearby. And my husband actually
was on a Jeep on the road with GPS deploying
sensors at the same time. These are two
earthquakes in the 90s that happened on the eastern
shore zone in California. These are showing
you the magnitude with time of the earthquakes. The black ones are, you
see the aftershocks. because they happen
after the main shock, shown with this cross. And the red ones
are the foreshocks. Do you see any pattern? No. There are some
earthquakes, but are they foreshocks, aftershocks of
the previous earthquakes? We don’t know. So these are kind
of the observations. In the lab though, if you were
to make earthquakes in the lab, the system is a lot
more predictable. And so these lab quakes,
they win every time. I’ll show you at the end. We have a nice plot on this. These lab quakes
behave so nicely. They behave like the model,
not like the Earth, apparently. But these lab quakes are showing
you the onset of seismicity. You can see those dots,
same kind of style here, getting more
and more frequent as we go, right
before the earthquake. And this is just showing
you kind of an acceleration of the step on the earthquakes. So in the lab, they behave
like we want them to be. In nature, it’s just
not, it’s not there yet. And there are two debated models
for how earthquakes start, and no one, I think, has clearly
stated this is the one model. I think it’s probably, as
always, a combination of both. . These are the
cascade model, where you have a bunch of
pops on the fault. They all pop and
trigger each other and then eventually becomes
a big one, in which case it’s very hard to
say which one is going to trigger the last one. Or you can have
this one, which is kind of a slow growth, a seismic
slip which accelerates and then becomes a big one. And this is rarely observed. And like this one. But it’s a lot more
consistent with models and lab experiments. And so there were many
papers this summer that discuss these two
models, and then we don’t have the data resolution
to actually validate those two. But this is the grail of
earthquake prediction. If you could actually see those
eventually and classify them as precursor, then
maybe you would be able to be before
the earthquake. So I would say that what
happened in nucleation, I think maybe we rarely
observe them, usually for large earthquakes. But we have observed some. So I think looking
back at the data, we may have hope to study
those phases a lot more. So now, the user, the
person, the surgeon, the eye surgeon with
his laser is saying, when should I stop the laser? When is the shaking coming
if there is an earthquake? And so, I’m not
skipping this phase, I’m just saying, who cares about
when the earthquake happened? I really care just about the
shaking and how much shaking. And this becomes
actually a different side of seismology, which is very
much oriented toward operation and user and, you know,
what matters for people is really just the amount of
shaking, whatever earthquake style we have. So, if these are your
science question, then we can look
at other things. So I want to focus the
rest of this discussion on what happened during the
rupture and right after, and how can we help the
early warning of the shaking. And this is– I’m an amateur at this, but
I wanted to show you briefly what the system of the
earthquake early warning is about. And I heard today
that the funding will be continued, which
is nice when you already spent 20 millions of
dollars, to have it finally an official earthquake early
warning system in the United States. Mexico has one since the 90s. So we’re developing– Scientists are developing
one in California, and we’ll see the
West Coast in general. And the system is as such. Phones and internet is
faster, propagates faster, than seismic waves. That’s where you win. That’s basically it. If you have an earthquake
that happens on a fault, some ways will propagate. The first one, the yellow one. The first one is not
the most damaging, but he carries a
lot of information. So what you want is having
sensors close by the epicenter. Then you have more to say well,
this is an earthquake and not a car driving by, so you may
need a few sensors for this. But then the information is no
longer sent with seismic waves, but with phones and
communication like this, so then the data gets into
the earthquake [INAUDIBLE],, and then there’s a
decision-making module which says, this will be the
size of the earthquake, therefore this will
be the shaking. Should we alert, yes, no. All of these modules are quite
complicated and elaborated. And then it gets to the people. And there are apps,
you can have it on your phone that
would tell you the level of shaking of
intensity of such will arrive at x seconds, or in x seconds. So the basis for the
earthquake early warning, it’s not to say the
earthquake will happen. It’s to say an
earthquake has happened, now the shaking will
happen in such time. So you can already see
that the further away you are from the fault, the greater
advance knowledge you may have. If you’re very
close to the faults, you may not have a
second of warning. So this is not even a point
almost of having warning. Although I’d like to say
that that one second, and I always think
about this eye surgeon with a laser in your eye,
and maybe that one second we’ll just shut off the
laser and that might actually save your eye. So even one second is important. In Japan, in the
nuclear power plant, they have seismometers
three kilometers down. And this three kilometers down
with a 3 kilometers per second, wave speed, gives you
one second warning, is sufficient to shut
down the operations. So even those few
seconds matter. But for the person, you and
I, or [INAUDIBLE] next week, sorry, you might want
some level of warning. And so a lot of
the questions have been if the earthquake
just happened, it may last tens of seconds,
maybe 30, 70 seconds. As soon as the
earthquake happens, I want to know how
big it is, I want to know how much shaking it is. Because the longer we
wait for that warning, the less Warning we’ll have. And so a lot of the questions
have been both in the– the use of data for
earthquake early warning, but in terms of
physics of earthquakes, is this beginning pulse,
this beginning onset, will carry the information
of the size of the event. Another way to say
this is we call that the determinism in earthquakes. Is the beginning
of the earthquake carries enough
information that we know what the end of the
earthquake is going to be like, therefore we can
tell what magnitude, therefore we can tell
what type of shaking. And so a lot of
the questions are– There’s a lot of
skepticism, again related to this Parkfield failed
experiment about earthquake determinism. And so in publications,
I was recommended to not use this term, it’s
politically not correct. But all it says is
there any information at the beginning that tells us
the evolution of earthquake? And so it’s just a word, and
I think there are sides to it, and I’m showing
you the both sides. And we’ll take one side, we’ll
see at the end of this talk. But basically there
are two camps, that’s why we have debates. They are the camp that say yes,
earthquakes are deterministic. If we look at the
onset of earthquakes, I know what the final
size of the earthquake will be before it even ends. And then you have the
no camp, which says look, all earthquakes look– [INAUDIBLE] this
started the same way. And I’ll explain those plots. But the bottom line
is that we have different data, different
ways to process it, different school of thinking. Those are probably the same
earthquakes and we’re just– we’re not coordinated enough
to be very systematic yet in our approaches to analyze the
data in a very objective way. So on this camp,
these are plots that show time of the earthquakes
to some normalization. This is in log scale of the
peak ground displacement. And the blue– The colors are showing you the
magnitude of the earthquake. And basically you see that
the way the displacement grows with time is not the
same for magnitude 9 or a magnitude 4 Others say– This is a 2014 study that said
small earthquakes start faster than big earthquakes. This study actually
shows the opposite, so I’m not sure the
validity of this. But still, it’s about
this growth, basically. This group shows
that in [INAUDIBLE].. By the way, in earth science,
we have so many scales that we put everything in
log time, spatial scales. We like to collapse things
into this lab log scale. So this is the log of the peak
ground displacement in meters. This is the log since
the P wave started. And so you can see
these are bandwidth with earthquake magnitudes. You have a lot of– I don’t think we’re really
good statistics, to be honest. Because we think that
everything is log normal, so we think median and means
and we try to respect this but it’s not necessarily robust. Nonetheless, if we take
the median of these curves, of P waves, which are the
first waves that arrive, you can see that the beginning,
whatever the magnitude you are at, the kind of
look one in the same. But this is one data
set, this is one team. These guys are saying, I’m using
another data, I don’t see this, and debate goes
on and on and on. And so, it’s really
interesting to think about there’s a lot of
diversity in earthquake and so really, how
should we study them? We have all these scales
and [INAUDIBLE] scales and temporal scales
and so, I think we’re not really
fully integrated yet, but I think there’s
hope to do that. I’m not yet fully
integrated, but I’m just raising the question and I
think eventually we should think about all the scales. And so– pardon for the
lack of resolution here, there’s one way to do it, which
is the camp that I just showed you earlier, is you say, I have
all these diverse earthquakes and I’m just going to
stack them together. I’m going to look at
the average earthquakes. And I was telling you
everything’s about lognormal. So what we do is a lot about
combining those messages and then trying to make this
average view of the earthquake. And so, if you actually
think this medians and median [INAUDIBLE] on all these
earthquake magnitude, this is time, this is
getting kind of like what the P wave would look like. These people found a
triangle function that makes it to science right away. Triangles are weird
in earthquakes. Everything should be lognormal
and should not have any kinks. But what they saw
is if we average everything over everything
then it looks like the same. So that’s saying there’s just a
common behavior of earthquakes. I showed in some of my
work that if I do the same, there’s some differences
but they’re only related to earthquake
depth not earthquake sizes. So in terms of
predicting magnitude you’re kind of out of luck. So that’s one side you can–
strategy for science or saying, take all the earthquakes,
average everything, find the commonality
between earthquakes, simplify the system
as much as possible. Or you can embrace
the complexity. Earthquakes are so
different from one another, and I was telling
you it’s lognormal. So there’s a lot of
differences of those scales. So this is just showing you
a little snippet of what data sets we’re working on is
how different earthquakes look like. And if they are so different,
what can we say about it? Is there any way we
can predict earthquakes or the size of the earthquake
if they all look different? Another example. We got thousands and thousands
of these functions that look like earthquakes, but
really they vary a lot. So part of the
research that I started doing this summer with a really,
really good team of students, I was so fortunate. I really love
working with students who have no a priori on
seismology whatsoever and start asking
these questions that are so fundamental,
without having the history and the package of why
they should not be saying determinism. So this term was really
very creative, not stubborn, but you know very determined
in solving the problem. In actually found a
really interesting pattern in the complexity
of earthquakes. So it’s is no longer about
let’s simplify everything into averaging everything. He said let’s look
at this complexity, embrace it and find patterns. And so I will
summarize it in such. We have an earthquake here. It’s a big earthquake, probably
a magnitude 7.5, maybe 8 given the duration here,
60 second of duration. This function could
be like a P wave, so you could see that
in a seismometer. I’m happy to talk more
details about that later. And I’m showing you two
types of earthquakes. We have the big one in orange,
the small one in turquoise, teal, beautiful color scheme. And they look different, right? So one is really
complex, lots of bumps. They’re not the simple
triangle function. If you look at
individual earthquakes, there are a lot
of richness in it. And then the small
one has one bump. And so he basically said,
let’s count the bumps. And he made bumps with
Gaussian functions, lognormal, we love it, and he
composed these functions by just adding and summing
all of these log functions and counting them as
what we called subevents. They are not a full
earthquake, they’re just a small earthquake,
a small event that belongs to the earthquake. And so it seems like
large earthquakes have a lot of bumps, subevents,
and small earthquakes have fewer bumps. So that’s increased complexity
with earthquake size, which is not that surprising. What was interesting is we could
model this very simply in-house with simple fracture
mechanics models, where we created earthquakes
with a set of parameters I can explain later. But basically what we found
is the big earthquakes that a lot of bumps and the
small earthquakes– smaller earthquake has one bump. OK. So the number of
bumps or some subevent grows with earthquake size. This x-axis here is
showing you the size of the main event,
the main shock, and these are the
number of subevents. You can see that with
growing size, this is Tohoku, and by all means, an outlier for
everybody to study earthquakes, but the biggest
one is the outlier. But if you look at ensemble
of the number of subevents, you have a growing
number of subevents. So big earthquakes
are more complex because they have
more subevents. However, it’s not we don’t
have a log scale here. And so actually, the
size of the subevents grows also with the
size of the earthquake. And that was really
the breakthrough, is that we don’t have a
characteristic bump that tells us a characteristic
length scale in the earth. Instead the earthquake
dynamics has this interesting and
intrinsic link scale that is characteristic
to the earthquake’s size. All of these dots
are a subevent that has a size shown in here
in log scale attached to the size of the main
earthquake shown here. These lines, this
is a one to one line saying the subevent
here has the same size as its master
earthquake, in which case this earthquake probably only
has one subevent to form it. This line here is showing
you that earthquakes would have about 100
bumps, 100 subevents, to form the earthquake. And so what you can see, it
is not an infinite amount. The subevents don’t
have the same size. These events actually grow
with earthquake magnitude. And so what was interesting,
the current scale is representing
the distribution. And so, these small
events tend to have mostly one main subevent,
versus the large event have a lot more. And if you were to
draw kind of a scaling, it’s not one to one but
some very strong scaling. So that means in terms of
structure of the earthquake, a small event has a small
bump and the big event has a big bump at the beginning. So let’s say you measure
this, you can tell crudely what size of the earthquake
it is, which is fine, the earthquake is done. But if you go to those
larger magnitudes, if you measure this
first bump, then you can predict or estimate the
final size of the event. Without the earthquake, the
earthquake is not done yet, you’re just making an inference
based on this organization of the seismic signal. And so we took this
independent data set, we looked at this recent
Indonesia earthquake in 2018, 7.5. It was large. Was not really on a
map fault, per se, it was not really on
the plate boundary. It’s one of those
earthquakes that just tell us we don’t know much. [INAUDIBLE] just on
plate boundaries. Earthquakes happen
in unforeseen places, so we can’t really just
predict an earthquake tomorrow, but these are the
earthquakes that remind us how young our field is. So the top function
is showing you the same, what will be a P
wave as a function of time. And then we fit in all
these Gaussian functions, joining these dashed lines. And at the bottom, this
is one of our attempts to say why it would be,
given this bump, our estimate of the magnitude of
the final earthquake. This green line is showing
you the past magnitude. This says, if I
integrate here this, I had a small
earthquake, shown here. And if I go on, a lot of
the earthquake has gone. And so by the end of
the day, the magnitude is whatever is gone until here. And so these dots
are showing you, with some kind of refinements,
that we do along time that our first estimate
here is within a few seconds of the earthquake. We are not far off
the final magnitude, although the earthquake
hasn’t really gone yet. And so, that was quite exciting. It’s not this
nucleation phase where we can see this aseismic slip. We don’t wait for the
end of the earthquake. We’re kind of like this nice
warning of about 40 seconds, possibly a warning ahead. So we’re close to a
minute more warning. And so to give you a sense of
how earthquake early warning works today, this is the one
is being estimated for Japan. JM is a Japanese
meteorological agency. These stands for earthquake
early warning system. Study from Minson and others. And they looked at a
few Japanese earthquake with those– STF stands for that function
I was showing you earlier. These are magnitude
of the main earthquake and these are time,
basically rupture time, of the earthquake. So these are small earthquakes. They are shorter. And these are the functions. I was showing you a green
one, this is a black one. This was a large
earthquake, magnitude 9. This was a magnitude
7, magnitude 6.9. And what they’re
showing you here is basically the estimate of the
earthquake early warning system there in Japan. So during the Tohoku earthquake,
they waited about 20 seconds after the earthquake started
to issue a first magnitude estimate. They soon realized that
the amplitude was big and so it was a
large earthquake. So they estimated within 25
seconds it was a magnitude 7. And then as time goes,
you get more data, you refine your estimate. Their final estimate
was really off by one or two of
magnitude, which is equivalent to a 30
times difference in size of earthquakes. And this is showing
you the function before, this is the true
size of the earthquake. So I was telling you we
failed at really predicting this Tohoku earthquake,
but if we were to look at other earthquakes. I’m just showing you dots of
where our prediction would be in time. And we really would be a lot
faster than current system if this were implemented
in earthquake. And so we looked at mini events. We have 3000 of those. And we made this figure that
shows relative to duration time of the earthquake. These are all those
subevents or bumps, estimated magnitude for the
last earthquake, the large one. And what we found is
usually between 10% and 20%, at the beginning
of the earthquake, we can tell what
magnitude it is. It has some nice features to it. I’m happy to talk about it. It has to do with the
structure of the subevents. But the bottom line is that
we can predict the magnitude before it stops. But it was not really
fancy, it was just truly finding a pattern in
the complexity of earthquakes. So these types of savings could
be between 10 and 30 seconds of saving of magnitude,
which is quite significant. So I wanted to show you a new
perspective on what machine learning, or buzzword being
artificial intelligence, could do for
predicting earthquakes. We’re going to turn back
into lab quakes, which predicts very
nicely already what earthquakes will look like. But I think there’s
a lot of hope to use those methods
to really try to find patterns in the
current data set that we have. And let’s say that
we currently work 10, 20 terabytes data set is normal,
especially for my post doc, who overloaded the
Harvard tester last night with many terabytes of data. So we have a lot
of data to look at. So this study is
showing you that they were able to predict when the
next earthquake, lab quake, would happen, just by looking
at these ambient vibrations. And so in lab, you can
replicate the earthquake cycle a lot faster, so you have
multiple earthquakes. And I was showing you the Alpine
fault it’s every 300 years. We don’t have 300
years of seismograms. And so you can kind of
speed up this process and look at those patterns
for the earthquake cycle. So what they found is looking
at these ambient vibrations is when the earthquake happened
here and the stress drops, the waves look like this. And then when the earthquake
is close to failure, you may have some of those
more impulsive events. But what was interesting
is they managed to using this
ambient time series to predict the time to
failure through time here during the experiment. And so their prediction
was shown in blue, with a random first algorithm. And then what the
experimental data was showing is shown in red. And these are the
validation part. And so it’s really stunning that
even though it’s not regular, these predictions were able
to show a weaker earthquake compared to a big earthquake. So I thought it was one of
the most promising study for machine learning
to predict earthquakes. In-house, our team
at Harvard has been able to predict the
location of aftershocks. And so as I was telling
you, we have the main shock and then we have many
aftershocks, which is something we forget about. If your building is already
shaken by the main one, you will have many
more earthquakes to go. The building is going to have to
deal with many more earthquakes and aftershocks. And so what this
study was showing is that if you look
at those black dots, the fault is showing at
these yellow patches. The red is showing basically
the change in stress, and then you can predict
from the slip that happened on the previous earthquake. And they worked with
multiple earthquakes here. And what they found is
that the traditional way of looking at where an
earthquake would be, given a change of
stress in the crust, failed to predict many,
many of the aftershocks. But instead, they found
within your network that you could tease
out a metric, a better metric to predict where
aftershocks would be, which has to do with
stress invariance. But basically these
approaches were successful at
predicting the location of the next earthquake. And so this is
short-term forecast. It’s not saying years ahead
we’ll have an earthquake. It’s saying we
had an earthquake, we know there will
be aftershock. And it’s saying where the
next aftershock could be. Other things that
machine learning has been really successful
at in seismology is detecting those very weak,
weak ground motions that are buried in seismic noise. And it has a really
good [INAUDIBLE],, that was one of the first to use
a deep neural network to detect earthquakes. And we all worked
in Oklahoma, we worked on the data set that
already had a few thousand events, but we were
able to detect something like 20 times more events. And these are shown here,
with those beautiful S wave and P waves all aligned. And we have hundreds and
hundreds of those events that we can look at the
wave forms and say ha, actually there
was an earthquake, we just had not detected it with
natural and more established methods. And so in terms of
prediction, I think there’s a lot of hope
with the new methods, and mostly the past
recording of data, where we can
potentially tease out some of these short-term
term forecasts. And I will stop here. Thank you. [APPLAUSE]

Leave a Reply

Your email address will not be published. Required fields are marked *