srl-2019245_supplement.pdf

Supplement

ary Material

Title

: Earthquake Early Warning ShakeAlert 2.0: Public Rollout

Authors

: M

onica

D. Kohler, Deb

orah E.

Smith, Jen

nifer

Andrews, Ang

ela

Chung, Renate

Hartog, Ivan Henson, Doug

las D.

Given, Robert de Groot, Ste

phen

Guiwits

This is the

upplemen

tary material

for

Earthquake Early Warning ShakeAlert 2.0: Public Rollout

The supplement include

s: 1)

two

additional

figures that are referenced

in the

main manuscript

(Figures S1 and S2); and

2) a

etailed description of ground motion

assessment tests

with two

figures that are referenced in this

supplementary

text

(Figures S3 and S4)

Figure S1

Schematic showing

ShakeAlert 2.0 algorithms EPIC and Finder, and the messaging

architecture illustrating similarit

ies, differences,

and message

sharing among the four

contributing regional networks

based

in Seattle, Pasadena, Menlo Park, and Berkeley.

Schematic shows data flow pathways starting from incoming seismograms (bottom), to alerts

sent to subscrib

ers (top). Different color po

lygon

s represent different purposes of each

algorithm and where they occur on the data

flow

pathway for each regional network. Light

blue boxes: waveform processors which process incoming raw waveforms, sy

mbolized by

seismogram image

, from the seismic stat

ions for FinDer. Violet boxes: waveform processors

which process incoming raw wa

veforms for EPIC. Gray circles:

ActiveMQ message

passing

software

instances

. Ma

genta boxes: FinDer, EPIC, and eqI

nfo2GM (

“

”

) algorithms. Dark

red boxes: Solution Aggregator a

lgorithm. Green boxes: Decision Module algorithm. Red

boxes: Heartbeat monitor.

See main text for algorithm details.

Large tan boxes indicate

different groupings of tasks where the top

most layer is referred to as the

‘

Alert Layer.

’

Line

segments indicate

data flow where arrowhead indicates that in some cases direction

is one

way, and in others

two

way.

‘

Subscribers

’

are the general public receiving the alerts. Note that

alerts are sent from three out of the four regional network locations.

Figure S2

. ShakeAlert 2.0 development, testing, and production environments and workflows.

Different color boxes represent the different purposes of each processor. From top to bottom,

Gray: processors where algorithms are developed or modified. Light blue

: source code

repository. Tan: environment where code is built, compiled where relevant, and linked to

libraries. Olive green: repository containing binary and

configuration files associated with

code. Brown: processors that are part of the testing environ

ment in

situ at the four contributing

regional network

operations

locations in Pasadena (PAS), Berkeley (BK), Menlo Park (MP),

and Seattle (SE). Dark green squares: pairs of Alert Production processors in

situ at three out

of the four contributing regional

network locations. Salmon squares: Pairs of Pre

alert

Production processors

situ

at each network location. Light green pentagons: Pairs of

processors

situ

at three out of the four network locations responsible for disseminating alerts

to general pub

lic. Curved lines: feedback loop between code and code test results. Solid lines:

one

way data

passing between

processors. Dashed lines: communication feedback loop

between c

ode developers and results of test

s performed on code. Tan oval

defines processors

and tasks that are formal components of ShakeAlert; the Development Environment

is, by

contrast,

under the control of the developers.

escription of

additional

ground motion

validation

tests

Validation of GMPE/GMICE implementation in

eqInfo2GM

validation of

GMPE/GMICE implementation is tested by

assess

ing

their use by

the

algorithm eqInfo2GM (Thakoor

et al

., 2019). Eqinfo2GM

is used to translate earthquake source

parameters to ground motion estimates and

contains

some critical differences

with respect

to the

ShakeMap software.

This analysis is carried out

to validate the GMPE and GMICE

implementations in eqInfoGM, and to understand the effect of deviations from ShakeMap

methodology; these include different treatme

nt of source

site differences, and

the effect of

voiding

source terms that cannot be determined (see Thakoor

et al

. [2019] for further details).

The

eqInfo2GM messages are generated for the source parameters

from

the ShakeAlert 2.0

DM alerts. This include

s both contour messages and gridded map messages. ShakeMaps are

then

also generated using similar latitude/longitude ranges

and GMPE/GMICE settings

, but no

observed station data are used. Next, a point

point comparison is generated for the historic

even

ts, comparing the ShakeMaps to eqInfo2GM maps

at the lower resolution of the latter

. Plots

are generated to compare SI (Shaking Intensity), PGA, and PGV for different regions, including

outhern California,

orthern California, Pacific Northwest, and

“Unknown” which corresponds

to different GMPE/GMICE selections. The comparisons between ShakeMaps and eqInfo2GM are

assessed from the union of the two grids

, looking at all earthquakes in the test suite, for grid points

with MMI II or greater

. The

average

SI MMI difference between

all points on the two grids

corresponding to

eqInfo2GM and ShakeMap

ranges

from

0.32 to 0.68, with a peak near 0.05.

Fig.

a shows the average SI MMI differences for each earthquake in the test suite. Each point in the

histogram

is an average over all points for every pair of eqInfo2GM

ShakeMap grids produced for

that earthquake’s first alert and its updates.

The Variance Reduction (VR) for

the

differences

ranges from 91% to 99.7%, where 100% is a perfect fit

(

Fig. S3

). Note

that VR is defined from

Thakoor

et al

(2019) as

푉푅

[

−

∑

푚푖푠푓푖푡

∑

표푏푠푒푟푣푎푡푖표푛

]

∗

100

(1)

originally based on Dreger and Woods (2002).

The

eqInfo2GM

algorithm

performs

similarly to ShakeMap in southern California and

northern California, and good matches

are observed

between eqInfo2GM and ShakeMap for the

historic test suite comparison. Looking at the comparison between ShakeAlert 2.0 and ShakeAlert

1.0 (our baseline), a

pproximately 64% of the comparisons have

an SI

VR of 99% or better, and

83% of the comparisons have

an SI

VR of 98% or better. With eqInfo2GM there are larger

variations for Pacific Northwest events in the historic test suite comparison. This is more

notic

eable for the PGA and PGV VR statistics than SI VR statistics because SI is based on the

logarithm of PGA and PGV values. Slight differences in the

way in which

vs30 data

are

used by

eqInfo2GM and ShakeMap calculations explain most of the discrepancies

; th

e extent of the vs30

map used by eqInfo2GM is smaller and has to be extrapolated into uncovered areas such as

Canada.

Examining the average SI difference for all regions, this test result comparison suggests

average SI errors of up to ~0.5 MMI units. The m

aximum SI difference, which looks at the

differences point by point (and includes small length

scale variability), is on average 1.0 MMI unit

(where the average is computed over all the point

point differences,

individually

for each

earthquake’s first a

lert and its updates)

, and as large as 2.5 MMI units

(

Fig. S3

. Larger

maximum

differences tend to occur for larger magnitudes, which may be due to how the source

receiver

distance is computed, first noted by Thakoor

et al

(2019). For M>5

Shakemap compu

tes a

Figure S3

. Histograms showing: (a) the average shaking intensity (SI) MMI differences between

eqInfo2GM and ShakeMaps, (b) the range of variance reductions for the different SIs, and (c)

the range of maximum SI MMI differences between eqInfo2GM

and ShakeMaps. The values

in (a), (b), and (c) are computed for each earthquake’s first alert and its updates.

‘median distance’

(the distance that produces the median ground motions of all the possible fault

orientations that pass through the hypocenter

; Worden and Wald

, 2016)

that addresses the

deficiency of using an epicentral distance for an extended rupture. In contrast, eqInfo2GM always

uses epicentral distance unless a FinDer fault is available, which is rare in the test examples as

early alerts us

ually do not incorporate the FinDer fault when the

point

source estimate of

magnitude i

s below

6.0. Where a line source is available, eqInfo2GM uses the Joyner

Boore

distance where relevant for the GMPE.

The ShakeMaps

produced

for the historic test suite only use the

point

source

information

regardless of magnitude

. Since eqInfo2GM uses the finite

fault source for

earthquakes

, some

differences can arise

between the ShakeMap and the eqInfo2GM map message output for the lar

ger

events in the historic test suite

In contrast

, for the large finite

fault scenario earthquakes

discussed

, the ShakeMaps are calculated using finite

fault information; hence, these are a better test of

the finite

fault capabilities of eqInfo2GM. I

n genera

l, eqInfo2GM matches better in southern and

orthern California. There is greater variation in the near

source regions for the Pacific Northwest

(note 0.2

resolution for eqInfo2GM and a finer 0.05

resolution for the ShakeMaps, the latter are

down

sampled during comparison).

We also tested eqInfo2GM against ShakeMaps for t

hree

large

magnitude finite

fault

scenario

events: an

7.3 Hayward fault scenario

7.9 Southern San Andreas fault scenario

, and an

9.34 Cascadia scenario

(

Fig. S4

) (see

Data and Resources for details on scenario data sources).

Note that these comparisons show how well eqInfo2GM can generate a ShakeMap

like product

from source parameters, but not how well the ShakeAlert system will predict the source parameters

for this ty

pe of event or the timeliness of the alerts

For this

test

we first regenerated the scenario

event ShakeMaps with expanded boundaries to better compare with eqInfo2GM output. We then

translated finite

fault information (already produced by the scenario da

tasets) to FinDer

style

XML messages

by reducing fault plane descriptions from the scenario to line

source descriptions

by taking the upper or lower edge, as this is the limit of the capability of eqInfo2GM input

. Next,

we generated eqInfo2GM map messages

using the FinDer

style XML messages as input, to

generate map output. Last we compared the (independently produced) ShakeMap and eqInfo2GM

ground motion estimates. Figs.

a,d

show the

difference

in shaking intensity (in MMI units)

between the ShakeMap

ground intensity estimates (Figs.

b,e

), and the eqInfo2GM ground

intensity estimates (Figs.

c,f

)

for the three scenario earthquakes

. The Hayward fault

7.3

scenario earthquake has a good SI VR of 99.6%, and the Southern San Andreas

7.9 scenario

earthquake also has a good SI VR of 99.2%. To model

the

9.34 Cascadia scenario

, we used two

different FinDer line sources, one with shallow slip

(5 km)

and one with deep slip

(26 to 35 km)

The shallow slip scenario generated

an SI

VR of 94.8% and the de

ep slip scenario generated

an SI

VR of 96.3%.

Validation

ShakeAlert ground motion map accuracy

The

assessment of

full

ShakeAlert ground motion

accuracy

nvolves

comparison

eqInfo2GM map predictions with

USGS

NEIC Shakemaps which include station

observations.

This comparison is a better measure of how ShakeAlert

2.0

will perform compared to observed

ground motions

whereas the previous comparison

tested how well eqInfo2GM replicates

Shakemap’s ground motion prediction equations (i.e. source information without station data).

This work is discussed in

Thakoor

et al

. (2019

) and

summarized here, but these types of tests

are an ongoing part of Shak

eAlert.

Thakoor

et al

. (2019) implemented the comparison of

eqInfo2GM output to NEIC Shakemaps for regions

with

MMI ≥

. In their study, they found that

Figure S4

. Ground motion test assessment for t

hree

scenario earthquakes, and comparison

between Sh

akeMap and eqInfo2GM output. (a) Hayward fault scenario comparison between

ShakeMap and eqInfo2GM

showing difference in shaking intensity in MMI

units

. (b)

Hayward fault scenario ShakeMap. (c) Hayward fault scenario eqInfo2GM output. (d) San

Andreas fault scenario comparison between ShakeMap and eqInfo2GM

showing difference

in shaking intensity in

MMI units

. (e) San Andreas fault scenario ShakeMap. (f)

San Andreas

fault scenario eqInfo2GM output.

(g) Ca

scadia scenario comparison between ShakeMap and

eqInfo2GM showing difference in shaking intensity in MMI units. (h) Cascadia scenario

ShakeMap. (i) Cascadia scenario eqInfo2GM output. In all three cases, t

he differences

between ShakeMap and eqInfo2GM output are very small, illustrating the effectiveness of

eqInfo2GM estimates.

97% of the maps had

an SI

VR > 50%, 76% of the maps had

an SI

VR > 80%, and 46% of the

maps had

an SI

VR > 90%. In terms of SI dif

ference, 14% of the maps had a mean SI difference

within 0.25 MMI units and 62% of the maps have a mean SI difference within 1.00 MMI units.

For

the

August 24, 2014

6.0

South Napa

earthquake

, they had a VR of 99.5% when comparing

output to Shakemaps (with same source and no station observations) vs. a VR of 62% when

comparing output to ShakeMaps (with

earthquake

catalog source

parameters

and with station

observations).

References

Dreger, D.,

and B. Woods (2002). Regional distance seismic moment tensors of nuclear

explosions,

Tectonophysics, 356

, 1

3, 139

156.

Thakoor, K., J. Andrews, E. Hauksson and T. H. Heaton (2019). From earthquake source

parameters to ground motion warnings near you: the

ShakeAlert earthquake information to

ground

motion

(eqInfo2GM)

method,

Seis.

Res.

Lett.,

(3),

1243

1257,

doi:10.1785/0220180245.

Worden, C. B. and D. J. Wald (2016). ShakeMap Manual Online: technical manual, user’s guide,

and software guide, U. S. Geolo

gical Survey, doi:10.5066/F7D21VPQ.