Você está na página 1de 38

Development and Implementation of

the NGA models for ground-motion


prediction

Dr Peter J. Stafford
Willis Research Fellow
RCUK Fellow / Lecturer in Modelling Engineering Risk
Imperial College London
p.stafford@imperial.ac.uk
Questions & Issues to be Addressed
What are the NGA models, why and how were they
developed and what role do they play in catastrophe
modelling?
The USGS routinely update National Seismic Hazard Maps.
The latest release paints a very different picture of the
hazard - how should you interpret this change?
The treatment of epistemic uncertainty by the USGS during
the revision of the National Seismic Hazard Maps
What are the practical consequences of implementing the
NGA models and how should these models modify your take
on Californian earthquake risk?

What are they? Why Now? Development Epistemic Uncertainty Implementation


What are the NGA models?

We are concerned with estimating the extent of direct


physical damage sustained by a portfolio during an
earthquake event
Empirical ground-motion models (the NGA models)
provide the link between the occurrence of earthquake
events to the intensity of shaking that is experienced by the
structures in our portfolios
We need to be able to predict degrees of damage that
should be experienced by typical structural systems when
exposed to some level of shaking

What are they? Why Now? Development Epistemic Uncertainty Implementation


Predicting Earthquake Damage
Hypothesise
an
earthquake
NGA MODELS

Predict the Characterise the


ground motion structure at the
at a site of site of interest
interest

Estimate
the degree
of damage NGA MODELS

What are they? Why Now? Development Epistemic Uncertainty Implementation


Equivalent SDOF Oscillators

MDOF system

Equivalent
Borzi et al. (2008)
SDOF system

What are they? Why Now? Development Epistemic Uncertainty Implementation


Period of Oscillator & Building Height
2 storeys  T ~ 0.3 seconds
3 storeys  T ~ 0.5 seconds
10 storeys  T ~ 1.4 seconds
20 storeys  T ~ 2.7 seconds
Number of
storeys
divided by 7
is a pretty
good rule for
California

Chopra & Goel (2000)

What are they? Why Now? Development Epistemic Uncertainty Implementation


Acceleration Response Spectra

A response spectrum is constructed


by plotting the peak response against
the period of the SDOF oscillator

The peak absolute response of the


mass of the simple system is
determined for each period

Structures are idealised by simple


SDOF oscillators

Accelerograms are input to simple


SDOF oscillators

What are they? Why Now? Development Epistemic Uncertainty Implementation


Pseudo-spectral Acceleration
The NGA models provide estimates of the maximum
pseudo-spectral acceleration that would be experienced by
a single-degree-of-freedom oscillator with a damping ratio
of 5% and a particular response period.
This acceleration can also be related to the peak
displacement that a structure would experience
These accelerations have a very strong correlation with the
direct physical damage experienced by structures during
earthquakes
Hence, if we know the spectral acceleration, we have a
good idea of how much damage might occur

What are they? Why Now? Development Epistemic Uncertainty Implementation


From Earthquakes to Ground-Motion
The NGA models take a number of input parameters that
describe an earthquake scenario and provide estimates of the
distribution of ground-motion values that should be expected.

For any earthquake scenario, consisting of the magnitude of the


earthquake, the distance between the site and the source, the
nature of the ground at the site, etc. a very large range of
ground motions can be observed

We refer to this random behaviour as Aleatory Variability

What are they? Why Now? Development Epistemic Uncertainty Implementation


Aleatory Variability of Ground Motion

FACTOR
of 10

Chiou & Youngs (2007) 2004 Parkfield Eq. - Bommer (2006)

What are they? Why Now? Development Epistemic Uncertainty Implementation


Aleatory Variability & Epistemic Uncertainty

What are they? Why Now? Development Epistemic Uncertainty Implementation


Describe the Attenuation with Distance
The models estimate the
decay in spectral amplitudes
that occurs as waves radiate
away from the earthquake
source. They are often
loosely referred to as
attenuation relations

However, they also model


the initial strength of the
motions at the source as
well as the amplification of
waves that normally occurs
just prior to reaching the
Boore & Atkinson (2007) surface
What are they? Why Now? Development Epistemic Uncertainty Implementation
Predict Expected Spectra for a Scenario
The models are developed by
considering individual spectral
ordinates, i.e. a single equation is
presented for the spectral acceleration
for a period of 0.2 seconds, while
another is presented for spectral
acceleration at a period of 0.6 seconds.

However, the predictions for numerous


response periods may be obtained for a
given scenario in order to develop the
expected response spectrum
corresponding to a scenario

What are they? Why Now? Development Epistemic Uncertainty Implementation


The Next Generation of Attenuation

The term Next Generation of Attenuation was coined in order to


reflect the fact that the models being developed as part of the NGA
project aimed to supersede a suite of models that had been developed
during the early to mid-1990s.

THE OLD THE NEW


1. Abrahamson & Silva (1997) 1. Abrahamson & Silva (2008)
2. Boore et al. (1997) 2. Boore & Atkinson (2008)
3. Campbell (1997), Campbell & 3. Campbell & Bozorgnia (2008)
Bozorgnia (2003)
4. Sadigh et al. (1993, 1997) 4. Chiou & Youngs (2008)

5. Idriss (1991, 1993) 5. Idriss (2008)


Power et al. (2008)

What are they? Why Now? Development Epistemic Uncertainty Implementation


Why upgrade?

The decade or so that had passed had seen a number of large earthquakes occur
in regions where dense networks of digital accelerograms were in place. The pool
of available data was therefore significantly enlarged

Model Records Events Mmin Mmax Rmin Rmax Tmin Tmax


Abrahamson & Silva (1997) 655 58 4.4 7.4 0.1 220 0.01 5
Boore et al. (1997) 112 14 5.3 7.7 0 109 0.1 2
Campbell (1997) 266 173 4.7 8.1 3 60 0.05 4
Campbell & Bozorgnia (2003) 443 36 4.7 7.7 2 60 0.05 4
Sadigh et al. (1993, 1997) 960 119 3.8 7.4 0.1 305 0.05 7.5
Idriss (1991, 1993) 572 30 4.6 7.4 1 100 0.03 5

Douglas (2003)

What are they? Why Now? Development Epistemic Uncertainty Implementation


Requirements for NGA Models
The NGA modellers were set the task of developing new models that are
applicable over much larger distance and period ranges than had
previously been considered.

Develop equations for peak ground acceleration (PGA), peak ground


velocity (PGV) and pseudo-spectral acceleration for periods ranging
between 0 and 10 seconds
Provide for predictions from events with magnitudes up to 8.5, for
strike-slip, and up to 8.0, for normal and reverse faulting events
Enable predictions out to distances of 200 kilometres from the source
Allow for common site classification schemes, including NEHRP

Users will extrapolate the models anyway, so we might as well have the
experts do it for them

What are they? Why Now? Development Epistemic Uncertainty Implementation


Model Development

The first step in the model development was to compile a database of


strong-motion accelerograms from a large number of earthquakes
worldwide.
The individual records came from numerous recording networks but
were all uniformly reprocessed so that they could be compared on a
consistent basis and to identify the usable period range.
In addition to gathering the records, a very large amount of metadata
associated with every record was compiled and rigorously checked.

Although there are databases around the world with many more
accelerograms (e.g. Japan, COSMOS), the PEER database is currently
unrivalled in terms of the consistency of the record processing and the
very high quality of the metadata associated with the records.

What are they? Why Now? Development Epistemic Uncertainty Implementation


The NGA Database
The compiled database contains 3551 multi-component
accelerograms from 173 shallow crustal earthquakes ranging
in magnitude from 4.2 to 7.9
Each record had associated metadata that included
magnitudes, source-site distances, site classes, as well as up
to a further 120 additional pieces of information regarding
the record and the earthquake scenario.

The model developers were free to select any subset of


these records, but had to justify any exclusions.
This ability to select different records is the first component
of epistemic uncertainty associated with the NGA models.

What are they? Why Now? Development Epistemic Uncertainty Implementation


Datasets used by developers
The freedom to select subsets of the overall dataset led to very different
portions of the database being used by the various model developers
In all cases the contribution of the 1999 Chi-Chi, Taiwan, earthquakes was
very significant
Developer Team Number of Number of Chi-Chi Chi-Chi % Chi-Chi
Records Earthquakes Mainshock Sequence

Abrahamson & Silva 2754 135 318 1258 46


Boore & Atkinson 1574 58 380 380 24
Campbell & Bozorgnia 1561 64 381 381 24
Chiou & Youngs 1950 125 208 813 42
Idriss 942 72 152 700 74
Abrahamson et al. (2008)

What are they? Why Now? Development Epistemic Uncertainty Implementation


Numerical Constraints

As the developers were asked to produce empirical models


to cover scenarios that were not well-supported by data, a
series of numerical analyses were undertaken in order to
provide constraint to the models:
Simulations of rock motions
One-dimensional site response analyses
Three-dimensional basin response analyses
Development of a directivity model based on isochrone
theory

What are they? Why Now? Development Epistemic Uncertainty Implementation


The final models
AS08 BA08 CB08 CY08 I08
Near-source saturation
    
Style-of-faulting
    
Rupture depth Through Rjb
  (RV) 
Hanging wall Through Rjb 
 
Shear wave velocity
    
Nonlinear site response Constrained Constrained Constrained N.A.

Soil/sediment depth Constrained Constrained N.A.

Magnitude-dependent   
Nonlinear Inter & Intra Intra event Inter & Intra
Abrahamson et al. (2008)

What are they? Why Now? Development Epistemic Uncertainty Implementation


The previous generation models
AS97 BJF97 CB03 Sadigh97 I93
Near-source saturation
    
Style-of-faulting
    
Rupture depth Through Rjb

Hanging wall
 
Shear wave velocity

Nonlinear site response

Soil/sediment depth

Magnitude-dependent   
Nonlinear 

What are they? Why Now? Development Epistemic Uncertainty Implementation


What are the actual differences?
Linear Site Response Higher resolution
Generic site classes no longer used characterise sites in terms of in predicitons
average shear-wave velocity (related to stiffness of the site)

Nonlinear Site Response Lower ground-


The degree of site amplification depends on the strength of shaking motions at short
source-site
Small amplification for strong shaking, large amplification for weak
distances
shaking

Surface-rupturing vs. Buried faults Lower ground-


motions from large
Buried ruptures cause larger ground-motions than surface rupturing
events magnitude events

Finite-fault Effects More complex


Near-field scenarios modelled more comprehensively
models, heavily
reliant upon Chi-
Chi
What are they? Why Now? Development Epistemic Uncertainty Implementation
Implementation within Logic Trees

Stafford et al. (2008)

What are they? Why Now? Development Epistemic Uncertainty Implementation


Implementation within Logic Trees

Bommer et al. (2009)

What are they? Why Now? Development Epistemic Uncertainty Implementation


Weak constraint at long periods
Not a major concern for most insurance purposes, but very important
for many other applications

30 storey
15

Bommer et al. (2009)

What are they? Why Now? Development Epistemic Uncertainty Implementation


Implementation within Logic Trees
The divergent predictions for many scenarios mean that loss estimates
may be highly sensitive to the ground-motion model that is used.
Common practice within probabilistic seismic hazard analysis is to
implement multiple ground-motion models within a logic tree
framework.

The idea is that by using multiple models we are able to account for
epistemic uncertainty associated with ground-motion prediction.

However, it is impossible, by definition, to know whether we are


capturing this uncertainty or not.
In reality all we are doing is obtaining multiple views through the use of
alternative, legitimate, models.

What are they? Why Now? Development Epistemic Uncertainty Implementation


Components of Epistemic Uncertainty
1. Uncertainty in selecting the
appropriate database
Addressed through the use
2. Uncertainty in selecting the of multiple models
appropriate model formulation

3. Uncertainty in estimating the


population mean with a finite May be addressed for
data set individual models through
statistical approaches, e.g.
4. Uncertainty in estimating the error propagation,
population mean due to bootstrap or jack knife
uncertainty in the predictor
sampling etc.
variables in the dataset
(magnitude, distance, Vs30, etc.)
Youngs (2006)

What are they? Why Now? Development Epistemic Uncertainty Implementation


USGS Logic Trees

To generate the National Seismic Hazard Maps, the USGS uses a logic
tree formulation during their probabilistic seismic hazard analysis

2002 USGS
2008 USGS
Four Models Equally Weighted
Three Models Equally Weighted
1. Abrahamson & Silva (1997)
1. Boore & Atkinson (2008)
2. Boore et al. (1997)
2. Campbell & Bozorgnia (2008)
3. Campbell & Bozorgnia (2003)
3. Chiou & Youngs (2008)
4. Sadigh et al. (1997)

Abrahamson & Silva (2008) NGA model not completed in time to be


incorporated into the 2008 revision of the hazard maps

What are they? Why Now? Development Epistemic Uncertainty Implementation


Capturing Epistemic Uncertainty
Impossible to know the extent to which epistemic uncertainty is being
accounted for
Use of similar datasets, and the same sources of numerical constraint, is
likely to suppress the true extent of this uncertainty

Too
similar?

What are they? Why Now? Development Epistemic Uncertainty Implementation


The USGS 2008 approach

Assume that the epistemic uncertainty associated with sampling and


parameter uncertainty for an individual model is 50% for scenarios
with magnitudes greater than 7 and distances less than 10km
50% relates to an additive factor of 0.4 on the natural logarithm of the
spectral acceleration
Compute additive factors dgnd for other magnitude-distance bins on
the basis of the relative number of earthquakes that have been
recorded in these bins
n = number of earthquakes with
n recordings in the M>7 R <10km bin
dgndi = 0.4 Ni = number of earthquakes with
Ni recordings in the ith magnitude-
distance bin
Model the epistemic uncertainty by incorporating additional branches
of the logic tree from each of the three adopted models

What are they? Why Now? Development Epistemic Uncertainty Implementation


Logic Tree used by USGS for Hazard Maps
Three NGA models allocated equal weighting
Additional model for epistemic uncertainty also incorporated
Double branch intended to account for all components of uncertainty

Petersen et al. (2008)

What are they? Why Now? Development Epistemic Uncertainty Implementation


Epistemic Uncertainty Model
n The same average dgnd terms are applied to all
dgndi = 0.4 three alternative models
Ni Petersen et al. (2008)

What are they? Why Now? Development Epistemic Uncertainty Implementation


Epistemic Uncertainty & the USGS

The approach adopted by the USGS is a novel one and is a meaningful


way of accounting for the uncertainty associated with predicting ground-
motions for scenarios where little empirical constraint exists

It is not perfect:
In theory, the dgnd terms cannot be the same for all models
The reference 50% uncertainty is a statistically based educated guess but
is impossible to validate or invalidate
The factors to adjust this 50% uncertainty for other magnitude-distance bins
are not perfect as they do not account for issues associated with nested
sampling

It is (probably) a good engineering solution to an intractable problem

What are they? Why Now? Development Epistemic Uncertainty Implementation


The Chi-Chi sequence
There is some concern that when the next big one comes the predictions
of the models may change again, e.g., is Chi-Chi too influential?

The NGA model developers took a considerable amount of care with the
records of the Chi-Chi event
The Chi-Chi event is a reverse-faulting event but this influence is taken care of
by the models
The motions are generally thought to be relatively weak for this event but this
effect is taken care of through random effects
The biggest influence will be upon constraining the attenuation with distance
and the near-source effects (not a major issue for California)

The incorporation of the Chi-Chi records is not solely responsible for the
reduction in hazard the reductions can all be explained by physics

What are they? Why Now? Development Epistemic Uncertainty Implementation


After the next big one...
NGA models account for many factors not previously incorporated

The developments have been facilitated by the occurrence of large, well-


recorded earthquakes mainly outside of the WUS some suspect these
have too much influence

The developments are also made possible through the re-evaluation of the
metadata for the records and the uniform processing

Catastrophe models will still be calibrated to reflect insured losses and


lower ground-motions may just be traded off with more damage or loss.

The differences in the hazard maps may be more severe than in the Cat
models

What are they? Why Now? Development Epistemic Uncertainty Implementation


Conclusions
The NGA models represent a significant advancement in the state of
empirical ground-motion modelling

The do not necessarily represent the best available science, but they
certainly represent the best available engineering

The approach of the USGS is an improvement upon previous efforts to


account for epistemic uncertainty but should be regarded as a work in
progress

It makes rational sense to regard the 2008 hazard maps as superseding


the 2002 versions

What are they? Why Now? Development Epistemic Uncertainty Implementation


Development and Implementation of
the NGA models for ground-motion
prediction

Dr Peter J. Stafford
Willis Research Fellow
RCUK Fellow / Lecturer in Modelling Engineering Risk
Imperial College London
p.stafford@imperial.ac.uk

Você também pode gostar