WEBVTT
Kind: captions
Language: en-US

00:00:02.899 --> 00:00:05.140
Good morning.
My name is Domniki Asimaki.

00:00:05.140 --> 00:00:09.360
I am a professor at Caltech.
And it is my great pleasure to

00:00:09.360 --> 00:00:13.049
be today here and share with you
the work we’ve been doing on

00:00:13.049 --> 00:00:16.730
developing a sediment velocity
model to improve the simulation

00:00:16.730 --> 00:00:24.250
of basin amplification effects
in southern California.

00:00:24.250 --> 00:00:28.440
We have put a lot of effort in
improving the algorithms and the

00:00:28.440 --> 00:00:33.570
computer software and hardware that
we use to simulate high frequencies

00:00:33.570 --> 00:00:37.180
and the propagation from the source
to the crust to the surface in original

00:00:37.180 --> 00:00:41.770
ground motion simulations.
The challenge now for the geotech

00:00:41.770 --> 00:00:48.330
engineers like myself is to find
ways to represent the fact that

00:00:48.330 --> 00:00:52.010
high frequencies can actually
see the near-surface soil.

00:00:52.010 --> 00:00:57.110
So these complex three-dimensional,
nonlinear heterogeneous anisotropic

00:00:57.110 --> 00:01:01.610
features of the shallow crust cannot
be captured anymore by correction

00:01:01.610 --> 00:01:05.860
factors if we strive to propagate
frequencies up to 10 hertz or more

00:01:05.860 --> 00:01:09.260
through the shallow crust.
And this is the same shallow crust

00:01:09.260 --> 00:01:12.710
that will create conditions or will
impose constraints through failure

00:01:12.710 --> 00:01:17.440
on very large ground motions that
we strive to simulate for very large

00:01:17.440 --> 00:01:21.030
earthquakes yet to come.
And so this shallow crust matters

00:01:21.030 --> 00:01:25.950
to engineers and will significantly
modify the very high frequencies

00:01:25.950 --> 00:01:28.600
we are trying to propagate
from source to surface.

00:01:28.600 --> 00:01:33.050
And this is – was the motivation
for the work I’m going to share.

00:01:33.050 --> 00:01:42.190
So the previous iteration of the –
of the shallow crust refinement in

00:01:42.190 --> 00:01:48.000
ground motion simulations of SCEC is
referred to as the geotechnical layer.

00:01:48.000 --> 00:01:51.930
This was a smooth – is a smooth
geometric representation of the shear

00:01:51.930 --> 00:01:59.000
wave velocity down to 350 meters,
which merges the Vs30 from geologic

00:01:59.000 --> 00:02:04.409
max linear surface to the
velocity models that 350 meters.

00:02:04.409 --> 00:02:10.479
And this merging had definite benefits
because we were able to refine

00:02:10.479 --> 00:02:13.560
the shallow layering and therefore
propagate high frequencies through

00:02:13.560 --> 00:02:16.209
a more refined presentation
of the shallow crust.

00:02:16.209 --> 00:02:22.049
But, by imposing constraint at
350 meters connection point

00:02:22.049 --> 00:02:27.139
between shallow layers and CVM-S,
this could mask impedance contrasts

00:02:27.139 --> 00:02:31.579
that lie somewhere in between
and also alter the basin geometry,

00:02:31.579 --> 00:02:35.310
particularly in the basin edges.
And so [inaudible] [inaudible]

00:02:35.310 --> 00:02:39.129
and others say – in the recent years,
have shown through simulations

00:02:39.129 --> 00:02:45.469
that this requirement – geometric
requirement or constraint deteriorated,

00:02:45.469 --> 00:02:48.840
in some cases, rather than improved
the ground surface predictions

00:02:48.840 --> 00:02:54.030
compared to the classic
model without using the layer.

00:02:54.030 --> 00:02:56.940
So that was the motivation for the –
for the development of the sediment

00:02:56.940 --> 00:03:00.529
velocity model. We wanted to
develop an idealized model based on

00:03:00.529 --> 00:03:06.060
observations. So also look at how
actual profiles, Quaternary sediments,

00:03:06.060 --> 00:03:11.779
vary in – deposits vary
in shear velocity with depth.

00:03:11.779 --> 00:03:18.620
And try to simulate – to capture,
in an idealized fashion, the natural

00:03:18.620 --> 00:03:23.340
deposits of shear wave velocity.
And we wanted it to be function of

00:03:23.340 --> 00:03:27.080
available physical properties,
so properties that are already

00:03:27.080 --> 00:03:32.919
embedded in the – in the velocity
models so that we would have a

00:03:32.919 --> 00:03:36.329
seamless integration of the velocity
models we develop with the existing

00:03:36.329 --> 00:03:39.900
community velocity model of SCEC.
And we also wanted to find rules

00:03:39.900 --> 00:03:44.069
to preserve the basin geometry
while refining the stratigraphy

00:03:44.069 --> 00:03:46.999
inside the
basin structures.

00:03:46.999 --> 00:03:54.980
The database that we used to perform
this work was adopted from a series

00:03:54.980 --> 00:04:00.489
of [inaudible] shown here.
We used about 1,000 different profiles.

00:04:00.489 --> 00:04:04.549
Merging those profiles into
a unified database was

00:04:04.549 --> 00:04:08.010
not always straightforward.
An example I’m going to show you

00:04:08.010 --> 00:04:13.999
here is from the near-surface structure
and how different methodologies –

00:04:13.999 --> 00:04:20.780
invasive versus non-invasive can
lead to schematic differences

00:04:20.780 --> 00:04:25.280
in the profile representation.
So these are four different examples

00:04:25.280 --> 00:04:32.220
from a narrowly defined Vs30 ranges.
And you can see that, overall,

00:04:32.220 --> 00:04:35.320
the comparison between the
different databases is very good

00:04:35.320 --> 00:04:38.660
within the same Vs30.
This is because we chose to

00:04:38.660 --> 00:04:43.569
stay within a certain geologic –
set of geologic observations.

00:04:43.569 --> 00:04:47.960
But, when we look at the near-surface,
then the non-invasive profiles,

00:04:47.960 --> 00:04:52.930
which are the black lines from the –
a profile database shows us the

00:04:52.930 --> 00:04:57.550
schematic reduction in the near-surface
because these are inverted profiles.

00:04:57.550 --> 00:05:01.370
The discretization is here, and the –
and the inversion algorithm allows for

00:05:01.370 --> 00:05:05.970
this reduction. Whereas, in the –
in the invasive cases, the first

00:05:05.970 --> 00:05:09.740
measurement is usually taken at
2.5 meters’ depth, and that explains

00:05:09.740 --> 00:05:14.849
why the shallow presentation of the –
of the near-surface layers is

00:05:14.849 --> 00:05:21.060
so uniform in the invasive cases.
So we found a way to map one set

00:05:21.060 --> 00:05:26.250
to the other by trying to understand
how the different methodologies

00:05:26.250 --> 00:05:30.310
of measurement affect the measurement
of the shear velocity profile.

00:05:30.310 --> 00:05:36.300
And choosing an average
variation with depth.

00:05:36.300 --> 00:05:40.680
Thankfully the differences were
only concentrated in the near-surface

00:05:40.680 --> 00:05:44.919
down to about 2, 2.5 meters,
which requires very low velocity

00:05:44.919 --> 00:05:48.740
[inaudible] hertz
frequencies can see it.

00:05:48.740 --> 00:05:53.050
We were able to merge all those
databases to one unified profile

00:05:53.050 --> 00:05:57.020
and then come up with an idealized
variation of shear velocity with depth

00:05:57.020 --> 00:06:03.469
based on [inaudible] [inaudible]
original scale input parameters.

00:06:03.469 --> 00:06:08.240
So the parameters were idealized
models Vs-zero, which is basically

00:06:08.240 --> 00:06:12.759
the average velocity of the
top 2.5 meters of the crust.

00:06:12.759 --> 00:06:19.039
And then this – multiply with k
and the exponent 1 over n.

00:06:19.039 --> 00:06:23.360
The range of the profiles has been
[inaudible] ranged from a Vs30

00:06:23.360 --> 00:06:29.740
between probably 100 meters
and 1,200 or more.

00:06:29.740 --> 00:06:35.400
We reliably found that we were
able to capture Vs30 up to 1,000 meters

00:06:35.400 --> 00:06:42.780
per second. The training model that
we used, we chose to split the shear

00:06:42.780 --> 00:06:46.860
velocity profiles to Vs30 bins
that were statistically equivalent.

00:06:46.860 --> 00:06:53.639
These are the empirical variation of the
mean and the standard deviation and

00:06:53.639 --> 00:07:00.120
the idealized fitting based on these
profiles here – this model here.

00:07:00.120 --> 00:07:03.979
The parameters Vs-zero, k, and n,
their variation with Vs30 shown

00:07:03.979 --> 00:07:09.590
in this location. The input parameter
for the velocity profile is Vs30.

00:07:09.590 --> 00:07:15.639
The decision of where the
sediment profile stops and the

00:07:15.639 --> 00:07:19.300
subsurface rock starts
is based on z-1,000.

00:07:19.300 --> 00:07:26.190
Data didn’t only show us how the
Vs30 varies at the near surface

00:07:26.190 --> 00:07:31.560
as a function of depth, but they
also showed us how we could –

00:07:31.560 --> 00:07:35.979
helped us understand how to randomize
the shallow layers. This is the variation

00:07:35.979 --> 00:07:42.580
of the thickness of the layer profiles
as mapped from the measurements.

00:07:42.580 --> 00:07:46.180
So, when we had the smoother
presentation of the velocity, we had to

00:07:46.180 --> 00:07:53.360
break it down into thinner layers
and then randomize each layer.

00:07:53.360 --> 00:07:56.169
The way that we broke it down
was on the basis of data.

00:07:56.169 --> 00:07:59.520
I’ll show you an example right away.
And, on the right-hand side, this is

00:07:59.520 --> 00:08:04.300
the region-specific, geology-specific
variation of z-1,000 with Vs30.

00:08:04.300 --> 00:08:11.550
This was very useful for cases where
the z-1,000 was shown to be zero.

00:08:11.550 --> 00:08:15.800
The site was reported in the Community
Velocity Model as being a rock site,

00:08:15.800 --> 00:08:18.970
whereas, the geology
suggested that Vs30 was low.

00:08:18.970 --> 00:08:25.009
In that case, we have to find a way to
marry those two – those two consistent

00:08:25.009 --> 00:08:27.960
[inaudible] presentations.
Then we define the z-1,000,

00:08:27.960 --> 00:08:31.919
a function of Vs30, and so, at that
location, we were able to then switch

00:08:31.919 --> 00:08:37.560
to the actual rock properties of
the Community Velocity Model.

00:08:37.560 --> 00:08:41.690
So here’s an example for a site
with 320 meters per second

00:08:41.690 --> 00:08:46.720
and a z-1,000 of 120.
You can see this is [inaudible] –

00:08:46.720 --> 00:08:52.090
the homogenized – the smooth
representation of the [inaudible] profile.

00:08:52.090 --> 00:08:56.750
The discretization is of the basis
of the data I showed you here.

00:08:56.750 --> 00:09:06.340
And the randomization is by choosing –
by selecting from the – from the

00:09:06.340 --> 00:09:16.370
variation of the parameters Vs-zero, k,
and n to perturb this profile according

00:09:16.370 --> 00:09:21.810
to the statistics of the actual profiles
we used to develop SVM.

00:09:21.810 --> 00:09:28.410
The way that this model worked in
implementation in a basin structure –

00:09:28.410 --> 00:09:33.430
an example is shown here.
This is a cross-section northeast

00:09:33.430 --> 00:09:41.911
of Long Beach. And you can see that –
you can see the smooth variation with

00:09:41.911 --> 00:09:47.130
depth but also the respecting of the
interface between soil and rock.

00:09:47.130 --> 00:09:53.180
And you can see here, this is a location
where z-1,000 was reported as zero.

00:09:53.180 --> 00:09:57.850
The velocity at depth was higher than –
I mean, at the surface was higher –

00:09:57.850 --> 00:10:01.630
the z-1,000 than
1 kilometer per second.

00:10:01.630 --> 00:10:08.310
But we were able to then merge the
Vs30 of the geologic feature with the

00:10:08.310 --> 00:10:15.580
z-1,000 of CVM by choosing
the empirical defined depth to

00:10:15.580 --> 00:10:19.340
1 kilometer per second,
filling that very shallow layer

00:10:19.340 --> 00:10:26.220
with SVM and then transitioning
to the stiffer subsurface rock.

00:10:26.220 --> 00:10:31.560
We validated SVM by comparing
the goodness of fit – by evaluating

00:10:31.560 --> 00:10:34.750
the goodness of fit of both
1D profiles and their corresponding

00:10:34.750 --> 00:10:37.820
amplification factors.
The comparison between –

00:10:37.820 --> 00:10:43.260
was between SVM, CVM-S –
so only the cluster the structure,

00:10:43.260 --> 00:10:47.230
and then CVM-H – have our model
with the geotechnical layer.

00:10:47.230 --> 00:10:51.020
And the validation was performed
on a hold-out data set of 43 profiles

00:10:51.020 --> 00:10:53.700
that were not used for
the training of our model.

00:10:53.700 --> 00:11:00.240
Two examples shown here.
This is a site where CVM-S had –

00:11:00.240 --> 00:11:06.140
was mapped as a – as a rock.
In that case, the geotechnical layer

00:11:06.140 --> 00:11:09.920
does a very good job of approximating
the shallow crust in model, and this is

00:11:09.920 --> 00:11:15.160
also reflected in the comparison
of the amplification factors.

00:11:15.160 --> 00:11:19.200
In the next example, SVM clearly
does a better job in capturing the

00:11:19.200 --> 00:11:25.130
measured profile compared to GTL.
But this is a case where CVM-S also

00:11:25.130 --> 00:11:28.250
had a variation of shear wave
velocity with depth in the –

00:11:28.250 --> 00:11:34.680
in a comparable Vs30 range.

00:11:34.680 --> 00:11:38.300
The ensemble of the comparison
between measurement –

00:11:38.300 --> 00:11:41.200
the simulation is shown here.
The definition of the goodness of fit

00:11:41.200 --> 00:11:46.250
that we gave was on the basis
of [inaudible] function.

00:11:46.250 --> 00:11:49.960
The measured profiles and the –
and the transfer function, you can –

00:11:49.960 --> 00:11:55.650
you can see here that the GTL spans
a range of goodness of fit

00:11:55.650 --> 00:12:02.640
from very poor to very good.
CVM-S stands to have a lot more sites

00:12:02.640 --> 00:12:06.860
that are sedimentary
deposits mapped as rock,

00:12:06.860 --> 00:12:12.191
which is why GTL tends to
improve this presentation.

00:12:12.191 --> 00:12:14.990
SVM is systematically
better for the –

00:12:14.990 --> 00:12:18.670
for the three hold-out
profiles of the data set.

00:12:18.670 --> 00:12:24.430
In transfer functions, the comparison
is less contrasting, particularly because

00:12:24.430 --> 00:12:28.380
the details of the near-surface layers
have are very critical for the

00:12:28.380 --> 00:12:32.171
high-frequency amplification.
And therefore, once you don’t capture

00:12:32.171 --> 00:12:36.690
those details, then the range of
goodness of fit for amplification

00:12:36.690 --> 00:12:42.660
factor of the three profiles is similar.
Nonetheless, both the mean and the

00:12:42.660 --> 00:12:47.150
scatter above the mean of SVM
was improved compared to

00:12:47.150 --> 00:12:49.460
the other two options.

00:12:49.460 --> 00:12:53.560
Currently, we have – in blue,
we have implemented the SVM in the

00:12:53.560 --> 00:12:59.100
Unified Community Velocity Model.
We are able to now generate a regional

00:12:59.100 --> 00:13:02.380
velocity model that include the
sediment velocity model

00:13:02.380 --> 00:13:06.140
in the shallow layers.
And we are now conducting

00:13:06.140 --> 00:13:11.910
a forward simulation using Hercules
at the [inaudible] Valley.

00:13:11.910 --> 00:13:15.760
The choice of the site was because
the location has – we have a lot of

00:13:15.760 --> 00:13:23.541
information about properties.
And you can see here the surface

00:13:23.541 --> 00:13:27.980
velocity profiles of CVM-S
with the geotechnical layer

00:13:27.980 --> 00:13:29.280
compared to
CVM-S with SVM.

00:13:29.280 --> 00:13:34.440
The location of the – of the basins is
similar, although you can see here,

00:13:34.440 --> 00:13:40.430
the ones we are – for example, this
location here is a contrast in the GTL,

00:13:40.430 --> 00:13:43.470
whereas, in the SVM, there is
a more smooth presentation.

00:13:43.470 --> 00:13:48.670
Preliminary results show that we
capture better the wave propagation,

00:13:48.670 --> 00:13:54.360
particularly in the basin edges,
compared to measurements of different

00:13:54.360 --> 00:14:01.940
events. But stay tuned because
the results will be published soon.

00:14:01.940 --> 00:14:06.570
Future work – we would like to
implement SVM at a site with a dense

00:14:06.570 --> 00:14:11.170
array so that potentially we can couple
the material property statistics with

00:14:11.170 --> 00:14:18.130
the spatial statistics of the variation
of the material properties.

00:14:18.130 --> 00:14:21.740
Correlation structures and
correlation length in X and Y.

00:14:21.740 --> 00:14:25.240
We would like to test it for a larger region,
so move from the

00:14:25.240 --> 00:14:29.010
Garner Valley, we scale
to a L.A. Basin scale.

00:14:29.010 --> 00:14:32.680
And then this framework that I showed
you might be specific for the geology

00:14:32.680 --> 00:14:36.870
of the sedimentary deposits in southern
California, but the framework is

00:14:36.870 --> 00:14:40.280
transferable, and the potentially
northern California would be a natural

00:14:40.280 --> 00:14:44.250
next step for us to look at the geologic
features, trying to cluster geologies

00:14:44.250 --> 00:14:47.690
together, and then understand
how the variation of the properties

00:14:47.690 --> 00:14:53.540
with depth lies in
each one of those features.

00:14:53.540 --> 00:14:57.190
The SVM is available in Python
routines and Jupyter notebooks

00:14:57.190 --> 00:15:03.700
at the – our GitHub site.
And a lot of this – of what I presented

00:15:03.700 --> 00:15:07.560
today can be found in this publication,
although there is [inaudible] with

00:15:07.560 --> 00:15:10.940
the implementation of SVM
and use CVM and [inaudible]

00:15:10.940 --> 00:15:11.820
the ground motion simulations.