WEBVTT
Kind: captions
Language: en-US

00:00:01.484 --> 00:00:03.736
[silence]

00:00:03.760 --> 00:00:06.080
Hello, everyone.
My name is Kevin Milner.

00:00:06.080 --> 00:00:10.056
I’m from the Southern California
Earthquake Center at USC.

00:00:10.080 --> 00:00:12.800
I’m going to be talking to you
guys today about operational

00:00:12.800 --> 00:00:16.000
earthquake forecasting
with UCERF3-ETAS.

00:00:16.000 --> 00:00:19.840
I’ll be sharing some lessons learned
from the 2019 Ridgecrest earthquake

00:00:19.840 --> 00:00:24.400
sequence and implications
for future UCERF models.

00:00:24.400 --> 00:00:26.800
This work is part of a broad
collaboration, and some of

00:00:26.800 --> 00:00:31.056
those collaborations are –
collaborators are listed here.

00:00:32.640 --> 00:00:35.360
Let me begin with
a tale of two earthquakes,

00:00:35.360 --> 00:00:39.816
each alike in magnitude –
let’s say magnitude 6.

00:00:39.840 --> 00:00:44.456
Now, the Reasenberg and Jones
model, which is a very useful model,

00:00:44.480 --> 00:00:48.640
forecasts an approximately
5% probability of each of these

00:00:48.640 --> 00:00:52.856
earthquakes triggering a subsequent
magnitude 6 or larger earthquake

00:00:52.880 --> 00:00:57.600
within one week. Well, what if I were
to tell you that one of those earthquakes

00:00:57.600 --> 00:01:01.680
was right on the San Andreas Fault
with a really high – or some other

00:01:01.680 --> 00:01:05.600
high-slip-rate, well-studied fault.
Another one was kind of

00:01:05.600 --> 00:01:08.000
out in the middle of –
in the middle of nowhere.

00:01:08.000 --> 00:01:11.280
Our intuition suggests that
aftershock probabilities are greater

00:01:11.280 --> 00:01:15.280
for earthquakes near active faults.
So maybe we might think that the

00:01:15.280 --> 00:01:18.240
one on the left would have a higher
probability of triggering something

00:01:18.240 --> 00:01:23.440
truly large than the one on the right.
Now, the UCERF3-ETAS model was

00:01:23.440 --> 00:01:30.436
constructed precisely to formalize this
sort of intuition into a usable model.

00:01:31.600 --> 00:01:37.520
Now, UCERF3-ETAS is an extension
of the Third Uniform California

00:01:37.520 --> 00:01:42.080
Earthquake Rupture Forecast,
or UCERF – that’s the time-independent

00:01:42.080 --> 00:01:47.592
model used by the National Seismic
Hazard Maps, building code, etc.

00:01:50.000 --> 00:01:52.960
Then that model is extended with
a time-department model –

00:01:52.960 --> 00:01:58.880
UCERF3-TD, which adds Reid’s
elastic rebound theory, where the

00:01:58.880 --> 00:02:04.696
probability on a fault is
dependent on the date of last event.

00:02:04.720 --> 00:02:08.320
So maximum probability gains here,
for example, on the southern

00:02:08.320 --> 00:02:12.880
San Andreas and also up in the
Hayward and Calaveras can get

00:02:12.880 --> 00:02:16.880
up to, you know, a factor of 2 
over the long-term model.

00:02:16.880 --> 00:02:22.887
And they also decrease in the
proximity of prior large earthquakes.

00:02:24.560 --> 00:02:28.880
Now, UCERF3-ETAS is kind of a
new model that extends both of those.

00:02:28.880 --> 00:02:33.440
It merges kind of a typical
spaciotemporal clustering model –

00:02:33.440 --> 00:02:39.416
ETAS from Ogata –
with the UCERF3 finite fault model.

00:02:39.440 --> 00:02:42.880
So the inputs to the model could
include a historical catalog or

00:02:42.880 --> 00:02:48.320
scenario events or kind of trigger
ruptures that we put into the model,

00:02:48.320 --> 00:02:52.609
either on or between
UCERF3 faults.

00:02:56.320 --> 00:02:59.280
Now, the outputs of this model
are synthetic catalogs –

00:02:59.280 --> 00:03:03.920
suites of many of them.
And, from those synthetic catalogs,

00:03:03.920 --> 00:03:08.400
we compute things like triggering
statistics – X percent probability

00:03:08.400 --> 00:03:12.240
of triggering a magnitude Y –
or even fault-specific ones,

00:03:12.240 --> 00:03:17.176
so X percent probability of triggering
a magnitude Y on fault Z.

00:03:17.200 --> 00:03:20.960
And also realistic scenario
aftershock catalogs.

00:03:20.960 --> 00:03:24.080
So examples of typical
and extreme sequences.

00:03:24.080 --> 00:03:29.600
This is an example for a Parkfield
magnitude 6 as a main shock.

00:03:29.600 --> 00:03:31.600
Kind of typically, there will
be some small aftershocks.

00:03:31.600 --> 00:03:35.120
But, every once in a while, when one of
those happens, you might get a really

00:03:35.120 --> 00:03:39.821
large sequence where, here, the
southern San Andreas was triggered.

00:03:42.240 --> 00:03:46.640
Now, we call this operational-ish
earthquake forecasting

00:03:46.640 --> 00:03:50.080
because it’s still somewhat
of a manual process.

00:03:50.080 --> 00:03:54.329
UCERF3-ETAS requires
high-performance computing.

00:03:55.360 --> 00:03:58.400
You need many synthetic
catalogs for stable results.

00:03:58.400 --> 00:04:02.456
So, for the Ridgecrest results
I’m going to show, we ran 100,000

00:04:02.480 --> 00:04:06.560
catalogs for each Ridgecrest forecast,
and each took about seven hours

00:04:06.560 --> 00:04:12.960
on 36 compute nodes – a total of
about 2,500 CPU hours for forecasts.

00:04:12.960 --> 00:04:15.976
So it’s a big model.

00:04:16.000 --> 00:04:19.360
It’s run on demand.
So what that meant is, when I felt

00:04:19.360 --> 00:04:24.000
shaking on July 4th in the
Ridgecrest 6.4, I got to work.

00:04:24.000 --> 00:04:29.280
I had simulations running about
33 minutes after the – after the

00:04:29.280 --> 00:04:33.096
main shock on kind of
our local cluster at USC.

00:04:33.120 --> 00:04:36.960
And the first results from kind of
a partial simulation were posted

00:04:36.960 --> 00:04:44.720
to response.scec.org at 11:39,
so just about an hour after the –

00:04:44.720 --> 00:04:50.160
after the magnitude 6.4.
Now, those first results gave

00:04:50.160 --> 00:04:55.760
a 2.8% chance of another 6.4
or larger in the next week and

00:04:55.760 --> 00:04:59.976
just a 0.4% chance
of a magnitude 7 or larger.

00:05:00.000 --> 00:05:03.576
Which is less than what you would
get from a Gutenberg-Richter

00:05:03.600 --> 00:05:09.840
extrapolation. The 6.4 occurred in an
area between mapped UCERF3 faults.

00:05:09.840 --> 00:05:14.480
And, in this areas, UCERF3 is
anti-characteristic, and that ends up

00:05:14.480 --> 00:05:20.400
being something that’s an important
consequence of the UCERF3

00:05:20.400 --> 00:05:22.240
framework for this
type of – type of model.

00:05:22.240 --> 00:05:28.880
Now, the USGS forecasted a 2.3%
chance of a magnitude 7 the next week.

00:05:28.880 --> 00:05:31.840
And that was using the Reasenberg
and Jones model as updated in

00:05:31.840 --> 00:05:37.040
Page et al. 2016, which assumes
a Gutenberg-Richter magnitude

00:05:37.040 --> 00:05:40.880
frequency distribution.
So this was one of our kind of

00:05:40.880 --> 00:05:46.960
first lesson learned, where UCERF3
is anti-characteristic in off-modeled

00:05:46.960 --> 00:05:51.280
fault areas, which leads to lower
probabilities than generic forecasts

00:05:51.280 --> 00:05:56.136
if a sequence is happening in
one of those off-fault areas.

00:05:56.160 --> 00:06:00.560
But, if we knew about those structures
a priori, that might not be

00:06:00.560 --> 00:06:04.320
anti-characteristic in the model.
So I think this is a place where

00:06:04.320 --> 00:06:08.696
we need to rethink things in
the next – in the next model.

00:06:08.720 --> 00:06:13.976
We could re-evaluate fault
characteristicness in UCERF4,

00:06:14.000 --> 00:06:17.656
both on- and off-fault.

00:06:17.680 --> 00:06:20.960
And potentially, we could add
a logic tree branch which forced

00:06:20.960 --> 00:06:27.040
off-fault areas to be Gutenberg-Richter
or even an inversion constraint that

00:06:27.040 --> 00:06:33.520
could encourage Gutenberg-Richter
scaling on faults, where possible,

00:06:33.520 --> 00:06:37.200
where the data did not require
characteristic distribution.

00:06:37.200 --> 00:06:41.360
Another potential thing that could
change the characteristicness

00:06:41.360 --> 00:06:46.080
on faults is more connectivity.
So, if you have more magnitudes –

00:06:46.080 --> 00:06:50.560
if a fault is able to participate in
earthquakes that rupture more faults,

00:06:50.560 --> 00:06:55.760
it can – some of it’s kind of –
slip rate can be accommodated

00:06:55.760 --> 00:06:59.760
by larger ruptures, which would
result in a less-characteristic

00:06:59.760 --> 00:07:03.215
distribution on
an individual fault.

00:07:05.360 --> 00:07:10.320
So, when the 7.1 occurred,
the primary developer, which is me,

00:07:10.320 --> 00:07:13.520
was at a Dodger game.
You can see the players

00:07:13.520 --> 00:07:18.313
didn’t quite notice it, but we
were excited up in the stands.

00:07:18.338 --> 00:07:21.760
And since this is still a somewhat
manual process, it requires someone

00:07:21.760 --> 00:07:26.080
to submit the jobs. I couldn’t 
at that moment, so I messaged

00:07:26.080 --> 00:07:30.160
my collaborators – Bill Savran,
who had attended a prior training

00:07:30.160 --> 00:07:33.896
session, and he submitted
the simulations for me.

00:07:33.920 --> 00:07:38.800
We were able to get results to the
California Earthquake Prediction

00:07:38.800 --> 00:07:43.280
Evaluation Council – CEPEC –
via Tom Jordan just in time

00:07:43.280 --> 00:07:47.336
for their 10:00 a.m. call
the next – the next morning.

00:07:47.360 --> 00:07:49.680
The results are shown
here on the bottom right.

00:07:49.680 --> 00:07:53.120
It’s kind of an anti –
still anti-characteristic 1% probability

00:07:53.120 --> 00:07:59.760
of another 7.1 or larger in the
next week. And 0.6% chance of

00:07:59.760 --> 00:08:06.080
triggering a 7 or larger on the Garlock,
which was a key concern for CEPEC.

00:08:07.680 --> 00:08:11.680
And that was using just a 
point-source representation of the –

00:08:11.680 --> 00:08:17.200
of the earthquake. We updated
the results later on that – on that day

00:08:17.200 --> 00:08:21.896
with a finite fault that my
collaborator Ned Field drew.

00:08:21.920 --> 00:08:25.920
And we noticed the inclusion of
the finite fault increased the

00:08:25.920 --> 00:08:30.640
Garlock probability of – through
a pretty large [inaudible],

00:08:30.640 --> 00:08:37.280
up from 0.6 to 1.7% of triggering a
Garlock magnitude 7 in one month.

00:08:37.280 --> 00:08:41.040
And that’s because this fault
surface here is standing

00:08:41.040 --> 00:08:44.588
close to the Garlock
Fault down here.

00:08:46.000 --> 00:08:51.840
Now, later in that week, on Thursday,
the 11th of July, I drew a new fault

00:08:51.840 --> 00:08:55.680
surface kind of using the extents
from some InSAR data that I had

00:08:55.680 --> 00:09:00.776
seen and aftershock locations.
And this new surface got even closer

00:09:00.800 --> 00:09:05.780
to the Garlock, which increased the
probability all the way up to 4.3%.

00:09:07.920 --> 00:09:14.320
So it became apparent that there’s a
sensitivity to the finite fault sources.

00:09:14.320 --> 00:09:18.480
So here, showing three different
models – point source, the first model,

00:09:18.480 --> 00:09:23.360
the second model, and basically,
as these surfaces got closer and closer

00:09:23.360 --> 00:09:28.160
to the Garlock, those probability
of any 7.1 and also triggering

00:09:28.160 --> 00:09:32.778
the Garlock here, in blue,
increased greatly.

00:09:33.680 --> 00:09:37.600
So, with that apparent sensitivity,
we sought a more authoritative

00:09:37.600 --> 00:09:42.696
source than just Ned or
myself drawing polygons.

00:09:42.720 --> 00:09:46.240
So some options we have are
ShakeMap and the teleseismic

00:09:46.240 --> 00:09:53.016
inversions posted on ComCat.
They both have finite fault surfaces.

00:09:53.040 --> 00:09:57.576
We decided to use ShakeMap
as our preferred source.

00:09:57.600 --> 00:10:02.560
Here’s the latest version, Version 14,
which is based on InSAR data,

00:10:02.560 --> 00:10:04.560
and this has been our
preferred model ever since.

00:10:04.560 --> 00:10:08.296
We can fetch it
directly from ComCat.

00:10:08.320 --> 00:10:11.440
They had some crude finite fault
surfaces available shortly after

00:10:11.440 --> 00:10:15.920
the event, but it should be noted
that this particular finite fault source

00:10:15.920 --> 00:10:20.742
took four days to be
posted in ShakeMap.

00:10:21.680 --> 00:10:25.120
The teleseismic inversion
source was available quickly –

00:10:25.120 --> 00:10:27.840
about three hours,
but was unsuitable for our –

00:10:27.840 --> 00:10:33.628
for our uses. It’s enormous and
cut across the Garlock Fault.

00:10:34.560 --> 00:10:39.336
We thought about potentially tapping
into finite fault estimates from

00:10:39.360 --> 00:10:43.360
earthquake early warning
for a more immediate estimate.

00:10:43.360 --> 00:10:50.855
This is – this is the FinDer model, which
estimates a finite fault source on the fly.

00:10:50.880 --> 00:10:54.216
So that’s something we can
potentially tap into to get

00:10:54.240 --> 00:10:57.789
good early finite
fault estimates.

00:10:58.720 --> 00:11:02.880
So another lesson learned, that those
fault trigger probabilities are really

00:11:02.880 --> 00:11:07.120
sensitive to finite rupture sources.
And, as we would expect,

00:11:07.120 --> 00:11:13.016
higher for surfaces that extend closer
to other known faults in this model.

00:11:13.040 --> 00:11:17.360
And so it’s important to know that early
results are likely to change as better

00:11:17.360 --> 00:11:21.600
finite rupture models are published.
And maybe we could think about

00:11:21.600 --> 00:11:28.856
adding some sort of measure of source
uncertainty to reduce this sensitivity.

00:11:28.880 --> 00:11:33.651
We also spent a lot of time through this
sequence improving model outputs.

00:11:34.560 --> 00:11:38.160
A lot of automatic plot generation
for different time periods.

00:11:38.160 --> 00:11:40.640
And this is a magnitude
probability plot.

00:11:40.640 --> 00:11:44.000
It shows kind of order of magnitude
increase of probability on the

00:11:44.000 --> 00:11:48.080
Garlock Fault as a function
of magnitude over the long term

00:11:48.080 --> 00:11:51.877
time-dependent and
time-independent models.

00:11:52.880 --> 00:11:57.200
Include kind of examples of individual
catalogs at various percentiles.

00:11:57.200 --> 00:12:03.440
This is a median catalog.
This is a 97.5th percentile catalog.

00:12:03.440 --> 00:12:07.600
And this is the most extreme one,
where it triggered a Garlock rupture,

00:12:07.600 --> 00:12:10.560
which triggered a magnitude 8
on the southern San Andreas –

00:12:10.560 --> 00:12:16.640
really extreme simulation, but a good
way of kind of showing the range

00:12:16.640 --> 00:12:18.720
of possibilities
in these catalogs.

00:12:18.720 --> 00:12:25.016
Could also be, then,
beneficial to some potential users.

00:12:25.040 --> 00:12:30.296
We also are now doing automatic
comparisons with ComCat.

00:12:30.320 --> 00:12:35.016
So we can kind of see how our
forecast is doing against the data.

00:12:35.040 --> 00:12:39.120
Also working within the Collaboratory
for the Study of Earthquake

00:12:39.120 --> 00:12:45.175
Predictability, CSEP, to kind of
get more formal testing automated.

00:12:45.760 --> 00:12:51.816
So there’s some other
automatically generated plots.

00:12:51.840 --> 00:12:54.080
You can see here,
this is the spatial distribution.

00:12:54.080 --> 00:12:57.680
This is another thing that shows
the importance of using a good

00:12:57.680 --> 00:13:02.216
finite fault model to
forecast the spatial distribution.

00:13:02.240 --> 00:13:06.560
If you use a point source, you get
a very kind of circular forecasted

00:13:06.560 --> 00:13:09.280
spatial distribution.
But, using the ShakeMap sources,

00:13:09.280 --> 00:13:14.720
we can – we estimated the –
our outputted spatial distribution

00:13:14.720 --> 00:13:19.840
matches the data,
which is in cyan here, pretty well.

00:13:22.880 --> 00:13:25.816
So that was another
important lesson learned.

00:13:25.840 --> 00:13:28.320
But I would say the main –
the main takeaway from this was

00:13:28.320 --> 00:13:32.160
that real events are really important
to learn model sensitivities

00:13:32.160 --> 00:13:36.776
and what tools are required
to build and evaluate forecasts.

00:13:36.800 --> 00:13:40.320
And I think one of the important ones
for the next round of national maps

00:13:40.320 --> 00:13:43.840
is going to be talking about and
thinking carefully about the degree of

00:13:43.840 --> 00:13:48.240
fault characteristicness in the models.
Because they clearly have

00:13:48.240 --> 00:13:54.536
a large impact on our short-term
earthquake forecasts.

00:13:54.560 --> 00:13:57.280
But, in general, I think that,
despite these sensitivities,

00:13:57.280 --> 00:14:01.680
we learned that UCERF3-ETAS
provides useful information about

00:14:01.680 --> 00:14:05.920
fault-triggering probabilities
and can produce realistic synthetic

00:14:05.920 --> 00:14:11.680
aftershock catalogs. And may be worth
the investment to fully operationalize.

00:14:11.680 --> 00:14:15.576
And, until then, we’ll
continue to run it on demand.

00:14:15.600 --> 00:14:18.560
If you’d like to learn more about this
model and our experience with the

00:14:18.560 --> 00:14:22.240
Ridgecrest earthquake, we published
a paper in Seismological

00:14:22.240 --> 00:14:31.735
Research Letters last year. Take a look,
and if we have time for any questions,

00:14:31.760 --> 00:14:36.000
I am available. Thank you very
much for your – for your attention.

00:14:38.242 --> 00:14:44.687
[silence]