1. Creating the Virtual Seismologist
Tom Heaton, Caltech
Masumi Yamada, Caltech
Georgia Cua, Swiss Seismological Service

2. Earthquake Alerting … a different kind of prediction
What if earthquakes were really slow, like the weather?
We could recognize that an earthquake is beginning and then broadcast information on its development … on the news.
“an earthquake on the San Andreas started yesterday. Seismologists warn that it may continue to strengthen into a great earthquake and they predict that severe shaking will hit later today.”

3. If the earthquake is fast, can we be faster?
Everything must be automated
Data analysis that a seismologist uses must be automated
Communications must be automated
Actions must be automated
Common sense decision making must be automated

4. How would the system work?
Seismographic Network computers provide estimates of the location, size, and reliability of events using data available at any instant … estimates are updated each second
Each user is continuously notified of updated information …. User’s computer estimates the distance of the event, and then calculates an arrival time, size, and uncertainty
An action is taken when the expected benefit of the action exceeds its cost
In the presence of uncertainty, false alarms must be expected and managed

5. What we need is a special seismologist
Someone who has good knowledge of seismology
Someone who has good judgment
Someone who works very, very fast
Someone who doesn’t sleep
We need a Virtual Seismologist

6. Virtual Seismologist (VS) method for seismic early warning
Bayesian approach to seismic early warning designed for regions with distributed seismic hazard/risk
Modeled on “back of the envelope” methods of human seismologists for examining waveform data
Shape of envelopes, relative frequency content
Robust analysis
Capacity to assimilate different types of information
Previously observed seismicity
State of health of seismic network
Known fault locations
Gutenberg-Richter recurrence relationship

7. Ground motion envelope: our definition
Full acceleration time history
Efficient data transmission
3 components each of
Acceleration, Velocity,
Displacement, of
9 samples per second
envelope definition– max.absolute value over 1-second window

8. Data set for learning the envelope characteristics Most data are from
TriNet, but many
larger records are
from COSMOS
70 events, 2 < M < 7.3, R < 200 km
Non-linear model estimation (inversion) to characterize waveform envelopes for these events
~30,000 time histories

9. Average Rock and Soil envelopes as functions of M, R
rms horizontal acceleration

10. horizontal acceleration ampl rel. to ave. rock site
Vertical P-wave acceleration ampl rel. to ave. rock site
horizontal velocity ampl rel. to ave. rock site
vertical P-wave velocity ampl rel. to ave. rock site

11. Estimating M from ratios of P-wave motions
P-wave frequency content scales
with M (Allen and Kanamori, 2003, Nakamura, 1988)
Find the linear combination of log(acc) and log(disp) that minimizes the variance within magnitude-based groups while maximizing separation between groups (eigenvalue problem)
Estimating M from Zad

12. Voronoi cells are nearest neighbor regions
SRN
STG
LLS
DLA
PLS
MLS
CPP
WLT
Voronoi cells are nearest neighbor regions
If the first arrival is at SRN, the event must be within SRN’s Voronoi cell
Green circles are seismicity in week prior to mainshock

13. What about Large Earthquakes with Long Ruptures?
Large events are infrequent, but they have potentially grave consequences
Large events potentially provide the largest warnings to heavily shaken regions
Point source characterizations are adequate for M<7, but long ruptures (e.g., 1906, 1857) require finite fault

14. Strategy to Handle Long Ruptures
Determine the rupture dimension by using high-frequencies to recognize which stations are near source
Determine the approximate slip (and therefore instantaneous magnitude) by using low-frequencies and evolving knowledge of rupture dimension
We are using Chi-Chi earthquake data to develop and test algorithms

15. We are experimenting with different Linear Discriminant analyses to distinguish near-field from far-field records

16. 10 seconds after origin
20 seconds after origin
Near-field
Far-field
Near-field
Far-field

17. 30 seconds after origin
40 seconds after origin
Near-field
Far-field
Near-field
Far-field

18. Strategy for acceleration envelopes
High-frequency energy is proportional to rupture area (Brune scaling)
Sum envelopes from 10-km patches

19. Sum of 9 point source envelopes
Vertical acceleration

20. Once rupture dimension is known
Obtain approximate slip from long-periods
Real-time GPS would be very helpful
Evolving moment magnitude useful for estimating probable rupture length
Magnitude critical for tsunami warning

21. Conclusions
Bayesian statistical framework allows integration of many types of information to produce most probable solution and error estimates
Waveform envelopes can be used for rapid and robust real-time analysis
Strategies to determine rupture dimension and slip look very promising
User decision making should be based on cost/benefit analysis
Need to carry out Bayesian approach from source estimation through user response. In particular, the Gutenberg-Richter recurrence relationship should be included in either the source estimation or user response.
If a user wants ensure that proper actions are taken during the “Big One”, false alarms must be tolerated

23. Distinguishing between P- and S-waves

24. 3 sec after initial P detection at SRN
Single station estimate:
Prior information:
Voronoi cells
Gutenberg-Richter
M, R estimates using 3 sec
observations at SRN
No prior information
8 km
M=4.4
Epi dist est=33 km
M=5.5
Note: star marks actual M, RSRN