Scipedia

Maeda, T.

Kohiyama, M.

Yamashita, T.

Accuracy improvement of damage classifier for a wooden building using long short-term memory with response surface method

Ryu, H.

Byun, J.

Edwards, M.

14th International Conference on Structural Safety and Reliability - ICOSSAR'25 
1-6 June, 2025 - Los Angeles California, USA 
       R. Ghanem, E. Johnson, S. Masri and A. Olivier (Eds) 
 
 
 
 
AN EXTENSION TO THE USGS SHAKECAST FOR SYSTEM-LEVEL IMPACT 
ASSESSMENT USING A BAYESIAN APPROACH 
HYEUK RYU¹, JI-EUN BYUN² AND MARK EDWARDS³ 
 ¹Geosciecne Australia 
GPO Box 378 Canberra ACT 2601 Australia 
Hyeuk.Ryu@ga.gov.au 
 
²	James Watt School of Engineering, University of Glasgow 
Oakfield Avenue Glasgow G12 8LT United Kingdom  
Ji-Eun.Byun@glasgow.ac.uk 
 
³Geosciecne Australia 
GPO Box 378 Canberra ACT 2601 Australia 
Mark.Edwards@ga.gov.au 
 
Key words: ShakeCast, Bayesian network, bridge, earthquake. 
Abstract. The United States Geological Survey (USGS) ShakeCast provides near-real-time 
post-earthquake impact and inspection metrics for critical infrastructure facilities following an 
event. The system sends automatic notifications with information on inspection priorities based 
on damage estimates using a suite of fragility models and ShakeMap ground motion parameters. 
The current system provides only asset-level damage estimates, which can be further used for 
system-level impact assessment through a Bayesian approach. This paper describes a 
methodology for an application of a Bayesian approach using the product of the ShakeCast and 
demonstrates an example of its application for the road bridge network located in the Wheatbelt 
Region of southwestern Western Australia, inland from Perth. This region has been the subject 
of detailed collaborative research on infrastructure assets that has led to an operational piloting 
of ShakeCast in Geoscience Australia’s National Earthquake Alerts Centre. The bridge network 
in this research consists of approximately 100 bridges with earthquake vulnerability attributed. 
The performance of the selected routes is assessed assuming the bridges are the only vulnerable 
elements in the road network that can fail due to ground shaking. This proposed extension to 
ShakeCast will enable road transport asset managers to assess the likelihood of system 
operational status near-real-time following an event. 
 
 
1 INTRODUCTION 
The United States Geological Survey (USGS) ShakeCast is an openly available near-real-
time post-earthquake information management system developed by USGS [1]. The outputs 
from this system have been widely used by public and private emergency services, 
infrastructure operators, and facility engineers [2]. The system sends automatic notification 
Hyeuk Ryu, Ji-Eun Byun and Mark Edwards 
2 
 
with the information on inspection priorities based on damage estimates using fragility models 
and ShakeMap [3] ground motion parameters after an event. 
The current system provides asset-level inspection priorities based on asset-level damage 
estimates. These can be further used for system-level impact assessment once the system 
reliability is modelled based on the reliability of individual components and how they interact. 
This paper describes an application of a Bayesian approach using the product of ShakeCast 
and demonstrates an example of its application for the road bridge network located in the 
Wheatbelt Region of southwestern Western Australia, inland from Perth. Two cases were 
considered for demonstration: 1) preferred paths were chosen before the event, 2) viable paths 
were identified after the event. For both cases, Bayesian inferences were carried out to compute 
the following probabilities: 1) the probability of a path functioning after the event, 2) the 
probability of system failure, and 3) the conditional probability of a bridge failure given system 
failure.  
2 USGS SHAKECAST 
The goal of the ShakeCast system is to provide near-real-time damage and inspection metrics 
to facilitate users’ inspection priorities and protocols [1]. Damage probabilities are computed 
for each facility using ShakeMap intensity, and inspection priorities are compiled in descending 
order of the damage probability. For instance, Figure 1 shows the summary report for the 
bridges located in the Wheatbelt Region of southwestern WA, inland from Perth after the event 
of ML 4.0 near Wyalkatchem, WA occurred on 27 March 2025 
(https://earthquakes.ga.gov.au/event/ga2025gagmzg). This region has been the subject of 
detailed collaborative research on infrastructure assets that has led to an operational piloting of 
ShakeCast in Geoscience Australia’s National Earthquake Alerts Centre [4]. Table 1 is an 
excerpt from the report, where asset-level shaking intensity and probabilities of damages are 
set out. Note that gray, green, yellow, orange, and red correspond to None, Slight, Moderate, 
Extensive, Complete damage state, respectively, and the damage probabilities are expressed in 
percent. Given the low ground motion intensity of this event, damage to bridge asset was not 
expected, nor subsequently reported from field inspection. 
 
Figure 1: ShakeCast report for the bridges after the event of ML 4.0 near Wyalkatchem, WA. 
ShakeCast Report
ga2025gagmzg - MLa 4.0 - Wyalkatchem, WA - v1
Assessed Inspection Priority Category
Total Facilities Evaluated: 213
High: 0
Medium-High: 0
Medium: 0
Low: 0
None: 213
Disclaimer
The assessment of ground shaking severity and resulting bridge damage is approximate with much
associated uncertainty in the process. The damage probability assessed is based on fragility
functions published in the HAZUS technical manual (see link below) for classes of bridge structures
found in the US that have been mapped to Western Australian bridges using available bridge asset
information. For these reasons the attribution of inspection priority should be taken as an indication
of which assets have been damaged and that the actual damage will be subsequently assessed
through physical inspection on site. No responsibility is taken for the accuracy of the information
provided.
HAZUS Technical Manual:
https://www.fema.gov/sites/default/files/documents/fema_hazus-earthquake-model-technical-manual-5-1.pdf
Page 1/7
Hyeuk Ryu, Ji-Eun Byun and Mark Edwards 
3 
 
Table 1: Asset-level damage report, an excerpt from the ShakeCast report for the bridges after the event of 
ML 4.0 near Wyalkatchem, WA. PSA10 represents the pseudo-spectral acceleration at 1.0 sec. 
Name Alert level 
Distance 
(km) Gray Green Yellow Orange Red 
PGA 
(g) MMI 
PSA10 
(g) 
277 Gray 100.38 100 0 0 0 0 0.04 1.4 0.01 
278 Gray 101.38 100 0 0 0 0 0.06 1.7 0.02 
279 Gray 101.71 100 0 0 0 0 0.03 1.2 0.01 
 
3 BAYESIAN NETWORK 
An infrastructure network can be modelled as a system of components connected in either 
parallel or series or as a combination of both. As system performance is conditioned on the state 
of each component, the network can be modelled as a Bayesian network, which is a probabilistic 
graphical model representing a set of variables and their conditional dependencies through a 
directed acyclic graph (DAG). For instance, a road network can be modelled as a system 
consisting of bridges as components, as depicted in Figure 2, assuming the bridges are the only 
vulnerable elements in the road network that can fail due to ground shaking. 
 
Figure 2: Bayesian network diagram for a road network consisting of N bridges and M paths.  
In a Bayesian network, the joint distribution of the random variables can be formulated as the 
product of the conditional distribution given the parent nodes:  
𝑃(𝑆, 𝑃!, … , 𝑃" , 𝐵!, … , 𝐵#) = 𝑃(𝑆|𝑃!, … , 𝑃") × ∏ 𝑃(𝑃$|𝐵!, … , 𝐵#)"$%! × ∏ 𝑃,𝐵&-#&%!       (1) 
In this study, Bayesian network was constructed using a python toolkit MBNpy 
(https://github.com/jieunbyun/MBNpy), which is a Python implementation of the algorithm of 
matrix-based Bayesian network using the conditional probability matrices [5]. Two cases were 
considered: 1) preferred paths were chosen before the event, 2) viable paths were identified 
after the event. For both cases, Bayesian inferences were carried out to compute the following 
probabilities: 1) survivability, the probability of a path functioning after the event, 𝑃(	𝑃$ = 1), 
2) the probability of system failure, 𝑃(	𝑆 = 0), and 3) the conditional probability of a bridge 
B1 B2 B3 BN-2 BN-1 BN
P1 P2 PM
Bridges
Paths
System S
…
…
Hyeuk Ryu, Ji-Eun Byun and Mark Edwards 
4 
 
failure given system failure, 𝑃(	𝐵$ = 0|𝑆 = 0). For the first case in which preferred paths were 
chosen before the event, these probabilities were estimated deterministically, while lower and 
upper bounds of the probabilities were estimated for the second case due to the incompleteness 
of the identified survival and failure rules (owing to the large number of bridges) under the 
branch and bound for reliability analysis of general coherent systems (BRC) algorithm [6]. 
The survivability, 𝑃(	𝑃$ = 1) was estimated by summing 𝑃(𝑆, 𝑃!, … , 𝑃" , 𝐵!, … , 𝐵#) over all 
variables except 𝑃$. Similarly, the probability of system failure, 𝑃(	𝑆 = 0) was computed by 
summing 𝑃(𝑆, 𝑃!, … , 𝑃" , 𝐵!, … , 𝐵#) over all variables except 𝑆. The conditional probability of 
a bridge failure given system failure, 𝑃(	𝐵$ = 0|𝑆 = 0) was computed: 
𝑃(	𝐵$ = 0|𝑆 = 0) =
'(	*!%+,		-%+)
'(-%+)
= '(	-%+|*!%+)×'(*!%+)
'(-%+)
   (2) 
The lower and upper bounds of the probability of system failure were estimated 𝑃1(	𝑆 = 0) and 
1 − 𝑃1(	𝑆 = 1) , respectively. 𝑃1(∙)  denotes the probability obtained from incomplete 
identification of rules by BRC [6]. The lower and upper bounds of the conditional probability 
were estimated '
1(*!%+,-%+)
'1(-%+)
 and !2'
1(*!%!,-%+)2'1(*!%!,-%!)2'1(*!%+,-%!)
!2'1(-%!)
, respectively. 
 
4 BAYESIAN INFERENCE FOR BRIDGE NETWORK IN WHEATBELT REGION 
Bayesian inference was applied to the road bridge network located in the Wheatbelt Region 
of southwestern WA, inland from Perth. The road network connecting 6 regional towns consists 
of approximately 100 bridges spatially distributed in Figure 3. Each bridge was assigned a 
HAZUS bridge vulnerability class [7] based on structural characteristics which included the 
level of seismic design and number of spans. 
The first case assumed that preferred paths connecting two regional towns were chosen 
before the event. For the trip from Wooroloo, WA to Merredin, WA, three paths were chosen 
as depicted in Figure 3. Path 1 and 2 have a substantial section in common from Wooroloo to 
Merredin, while Path 3 goes through York instead of Northam. 
 
Figure 3: Map of the road bridge network located in the Wheatbelt Region of southwestern WA. 
Hyeuk Ryu, Ji-Eun Byun and Mark Edwards 
5 
 
For the scenario, a ground motion field was generated using the OpenQuake v3.20 [8] for 
Mw 6.58 event located approximately 20km from Northam, WA. Damage probabilities of four 
damage states were computed using the fragility functions associated with the HAZUS bridge 
class, which are implemented in the ShakeCast. It was assumed that bridge function failed if it 
sustained extensive or worse damage. Only three bridges of the selected three paths were 
computed to have the failure probability larger than 0.1 given the scenario, as marked in pink 
dots in Figure 3. 
Firstly, survivability, the probability of a path functioning after the event was computed and 
summarised in Table 2. The probability of system failure was computed to be 10.5% in which 
all paths would have unusable bridges. Optimal path decision can be informed by the 
combination of travel time and survivability. Secondly, the conditional probability of a bridge 
failure given system failure was computed and presented in Figure 4. This can be used, along 
with the ShakeCast assessment of probable bridge damage state, in prioritising recovery work 
across the bridges. Asset manager would have both an assessment of bridge damage and the 
likelihood that this outcome has impacted the connectivity of the transport system. 
Table 2: Probability of path functioning after the event for the pre-selected paths.  
Path 1 2 3 
Travel time (mins) 136.4 136.6 141.4 
Survivability (%) 59.8 77.1 54.1 
  
 
Figure 4: Probability of bridge failure given system failure for the first case. 
In the second case, a path was assumed to be disconnected if the travel time was longer than 
the threshold time, which was set to be 1.5 hours addition to the shortest travel time before the 
event. In this example, the maximum number of rules was set to be 20, which resulted in 8 
connected paths and 12 disconnected paths under the BRC algorithm. The shortest 5 paths are 
set out in Table 3 along with corresponding travel time and survivability. Note that the path 
number in the table was assigned by the order of travel time, which does not match to the 
number in the first case. The lower and upper bounds of the system failure were estimated to 
be 0.1% and 0.2%, respectively. The lower and upper bounds of the conditional probability of 
Hyeuk Ryu, Ji-Eun Byun and Mark Edwards 
6 
 
a bridge failure given system failure were computed and presented in Figure 5. 
Another simulation was carried out to illustrate limited path availability due to vehicle 
weight. For this each road section was assigned a vehicle weight limit along with road bridges 
on the section based on road type. Three values of the vehicle weight limit (50t, 75t and 100t) 
were assigned and paths were searched assuming a 75t vehicle could only pass along a section 
having a capacity equal to or larger than the vehicle weight. The simulation resulted in 4 
connected paths and 16 disconnected paths, and lower and upper bounds of the system failure 
estimated to be 8.5% and 10.2%, respectively. As expected, considering the vehicle weight limit 
resulted in higher system failure and a smaller number of viable paths. The available 4 paths 
are set out in Table 3 along with corresponding travel time and survivability. The lower and 
upper bounds of the conditional probability of a bridge failure given system failure were 
computed and presented in Figure 5. 
Table 3: Probability of path functioning after the event for the second case.  
Vehicle weight 
limit Path 1 2 3 4 5 
Not considered Travel time (mins) 136.6 140.3 141.4 149.0 154.3 Survivability (%) 77.1 72.3 54.1 49.8 54.4 
Considered Travel time (mins) 136.6 146.7 166.2 183.7 Survivability (%) 77.1 66.6 41.2 15.6 
 
 
Figure 5: Probability of bridge failure given system failure for the second case: a) vehicle weight limit not 
considered, b) vehicle weight limit considered.  
4 CONCLUSIONS 
In this research, a Bayesian approach was applied to extend the functionality of the USGS 
ShakeCast, by enabling the system-level impact to be assessed in addition to asset-level damage 
estimates. Through the example of the road bridge network located in the Wheatbelt Region of 
southwestern WA, feasibility of the extended functionality was demonstrated. The reported 
work has focused on a single event. The potential of this Bayesian approach will be further 
explored outside of an operational approach to examine how the Bayesian outcomes of a 
stochastic event set of earthquake impacts can provide information for prioritising bridge asset 
mitigation/replacement. 
 
Hyeuk Ryu, Ji-Eun Byun and Mark Edwards 
7 
 
REFERENCES 
[1] Wald, D. J, Lin, K, Kircher, C. A., Jaiswal, K., Turner, L, Luco, N., and Slosky, D. (2017). 
Development and Utilization of USGS ShakeCast for Rapid Post-Earthquake Assessment 
of Critical Facilities and Infrastructure, Paper No. 608, Proc. Sixteenth World Conf. 
Earthquake Eng. (16WCEE), Santiago, Chile.  
[2] Turner, L., Wald, D. J., Lin, K., Chiou, B., and Slosky, D.  (2018). State transportation 
agencies partner to deploy and enhance ShakeCast. Proceedings of the 11th National 
Conference in Earthquake Engineering, Earthquake Engineering Research Institute, Los 
Angeles, CA.  
[3] Wald, D. J., Quitoriano, V., Heaton, T. H., Kanamori, H., Scrivner, C. W., and Worden, C. 
B. (1999). TriNet "ShakeMaps": Rapid generation of peak ground-motion and intensity 
maps for earthquakes in southern California, Earthq. Spectra 15, 537-556, doi: 
10.1193/1.1586057. 
[4] Ryu, H., Edwards, M., and Carapetis A. (2024). Application of USGS ShakeCast for Rapid 
Post- Earthquake Assessment of Critical Infrastructure Facilities in Western Australia, 
Australian Earthquake Engineering Society 2024 National Conference, 21-23 November 
2024 Adelaide, South Australia, Australia 
[5] Byun, J.E., and Song, J. (2021) Generalized matrix-based bayesian network for multi-state 
systems, Reliability Engineering & System Safety 211, 107468. 
[6] Byun J.E., Ryu, H., and Straub, D. Branch-and-bound algorithm for efficient reliability 
analysis of general coherent systems, Reliability Engineering & System Safety, In Review. 
[7] FEMA (2022). Hazus Earthquake Model Technical Manual: Hazus 5.1, available at 
https://www.fema.gov/sites/default/files/documents/fema_hazus-earthquake-model- 
technical-manual-5-1.pdf  
[8] Pagani, M., D. Monelli, G. Weatherill, L. Danciu, H. Crowley, V. Silva, P. Henshaw, R. 
Butler, M. Nastasi, L. Panzeri, M. Simionato, and D. Vigano (2014). OpenQuake Engine: 
an open hazard (and risk) software for the Global Earthquake Model, Seismol. Res. Lett. 
85, 692–702, doi: 10.1785/0220130087. 


An extension to the USGS ShakeCast for system-level impact assessment using a Bayesian approach

Aijaz, N.

Ghosh, J.

Seismic response assessment of deteriorating highway bridges under earthquake induced landslides

Agrawal, S.

Hundia, N.

Liu, Z.

Bensi, M.

Deep learning motivated data imputation of tropical cyclone radius of maximum winds

Rodrigues, I.

Spence, S.

Beck, A.

Addressing statistical uncertainty in a surrogate model framework for Performance-Based Risk Optimization of structures subjected to seismic actions

Jiang, J.

Integrated performance-based wind collapse assessment of structural and envelope systems of engineered buildings

Sterrenberg, A.

Andriotis, C.

Stoter, J.

Modelling stochastic degradation and maintenance effects for the road network of Amsterdam: a multi-attribute data-driven approach

Birrell, M.

Astroza, R.

Li, Y.

A surrogate model approach for uncertainty treatment in performance-based engineering

Chojnicka, P.

Iannacone, L.

Limongelli, M.

Review of Structural Health Monitoring methods for the assessment of probabilistic models of imposed loads

Arcieri, G.

Duthé, G.

Muller, C.

Haener, D.

Papakonstantinou, K.

Straub, D.

Chatzi, E.

Graph-based multi-agent reinforcement learning for railway infrastructure decision support

Not a member yet