Deep learning-based retrofitting and seismic risk assessment of road

networks

Rodrigo Silva-Lopez1 , Jack W. Baker2 , and Alan Poulos3

1 Ph.D. Candidate, Dept. of Civil and Environmental Engineering, Stanford University, Stanford,
94305, California. Email: [email protected]
2 Professor, Dept. of Civil and Environmental Engineering, Stanford University, Stanford, 94305,

California
3 Ph.D. Candidate, Dept. of Civil and Environmental Engineering, Stanford University, Stanford,

94305, California

ABSTRACT
Seismic risk assessment of road systems involves computationally expensive traffic simulations
to evaluate the performance of the system. To accelerate this process, this paper develops a
neural network surrogate model that allows rapid and accurate estimation of changes in traffic
performance metrics due to bridge damage. Some of the methodological aspects explored when
calibrating this neural network are defining sampling protocols, selecting hyperparameters, and
evaluating practical considerations of the model. In addition to the neural network, a modified
version of the Local Interpretable Model-Agnostic Explanation (LIME) is proposed as a retrofitting
strategy that minimizes earthquakes’ impact on the system. The modified version (LIME-TI)
uses Traffic Impacts and rates of occurrence to aggregate the importance of individual damage
realizations during the computation of variable importance. This study uses the San Francisco
Bay Area road network as a testbed. As a conclusion of this study, the neural network accurately
predicts the system’s performance while taking five orders of magnitude less time to compute
traffic metrics, allowing decision-makers to evaluate the impact of retrofitting bridges in the system
quickly. Moreover, the proposed LIME-TI metric is superior to others (such as traffic volume or
vulnerability) in identifying bridges whose retrofit effectively improves network performance.

INTRODUCTION
Road networks are fundamental for the normal functioning of communities, allowing the flow
of goods and people throughout different geographical locations. Unfortunately, road networks
elements, such as bridges and roads, are vulnerable to damage from natural disasters such as
earthquakes. In the face of this challenge, decision-makers are confronted with using scarce
resources to minimize the consequences of disruptive events, improving the ability of communities
to recover from them.
In addition to being vulnerable, distributed systems, such as road networks, comprise thousands
of elements that interact with each other in a highly complex manner. Moreover, given that the
occurrence and characteristics of future events have large uncertainties, using a single scenario for
risk analysis is inadequate (Bommer et al. 2000), and properly informed decision-making requires
simulating the impact of several scenarios. Considering the complexities of evaluating the response
of the system to a single scenario, which must be repeated several times, the computational time

1
required to analyze distributed systems can impede decision-makers from evaluating actions to
minimize risk. One common approach to decrease these computational challenges has been to
select subsets of scenarios (Miller and Baker 2015; Han and Davidson 2012). However, the use of
fewer scenarios has implied an increase in the uncertainty of the predictions.
Motivated by the previous problems and the limitations of existing approaches, the first goal
of this study is to propose a neural network surrogate model to minimize computational times
required to analyze complex road networks. The focus of the proposed neural network is to help
decision-makers evaluate the effectiveness of different strategies to retrofit vulnerable bridges in
seismic regions.
As the second goal of this study, taking advantage of the neural network structure, a bridge
retrofitting strategy is proposed. This strategy uses a modified Local Interpretative Model-Agnostic
Explanation considering Traffic Importance (LIME-TI) algorithm as a variable importance model
to detect bridges in the network that contribute significantly to disrupting traffic.
The main contribution of this study is to prove that the use of neural networks allows rapid
and accurate surrogate models that can be instrumental for optimal decision-making for seismic
risk management of road networks. In addition, this work proposes a modified LIME-TI variable
importance algorithm that efficiently decreases the potential impacts of earthquakes.

RELATED WORK
Retrofitting bridges is the most common strategy used by decision-makers as a preparedness
policy against earthquakes. As limited resources are assigned to invest in the improvement of road
networks, the optimal allocation has been the motivation of several previous studies. One popular
way to solve this optimization problem is to use a two-staged stochastic optimization (Gomez and
Baker 2019; Miller-Hooks et al. 2012; Liu et al. 2009). Although these studies consider the best way
to retrofit bridges in road networks, the underlying computational costs in assessing these systems
added to the costs of performing the proposed optimizations limit the versatility of these approaches
and their ability to quantify the uncertainty in their results. Another popular technique used to
identify bridges for retrofitting is to rank them according to an importance metric. This ranking
system is the current approach used by the California Department of Transportation (Caltrans)
(Caltrans 2019). Other studies that rank bridges are the ones developed by Rokneddin et al. (2013),
and Basoz and Kiremidjian (1995). Using a ranking heuristic to retrofit bridges is intuitive and used
in practice; however, it limits the ability of the selected bridge set to reflect complex characteristics
of the road network, such as the interdependence of its components.
In order to evaluate retrofitting policies, decision-makers need to quantify the seismic risk of
road networks. This process is commonly performed by using Monte Carlo simulations in which
traffic performance of the network is obtained for several realizations of bridge damage, computed
from a set of seismic scenarios (Bommer et al. 2002). A limitation of these Monte Carlo methods is
that computing traffic performance is computationally challenging; therefore, the number of seismic
scenarios included in the seismic risk analysis is limited. Motivated by this, Han and Davidson
(2012) proposed a methodology to select a limited set of earthquake scenarios combining sampling
importance and optimization techniques. This concept was extended by Miller and Baker (2015)
to consider a proxy for damage as part of the optimization, and by Tomar and Burton (2021) to use
an active learning method for selection. Another technique to speed the road network seismic risk
assessment is to develop surrogate models for which traffic calculations are efficient. For instance,
Özdamar and Demir (2012) and Gomez et al. (2013) use hierarchical techniques to change the

2
resolution of the network while keeping consistency in relevant properties. Lim et al. (2015) use
a spectral clustering algorithm to develop a surrogate of the original network, computing zones of
greater importance connected by superlinks.
Deep learning techniques have become popular throughout earthquake engineering (Xie et al.
2020) and other areas of civil engineering. Previous efforts have used artificial intelligence, and
deep learning techniques tools to quantify road network seismic risk in an accelerated way (Nabian
and Meidani 2018b; Elhag and Wang 2007). Different studies have trained trained neural networks
to rapidly predict complex traffic metrics such as network connectivity or vehicle travel speeds
(Nabian and Meidani 2018a; Yu et al. 2017). Moreover, Alemzadeh et al. (2020) used neural
networks to optimize recovery of interdependent distributed systems, showing the utility of this
kind of tool for infrastructure system management.
Calibrated neural networks can also inform what drives road network performance. However,
the internal mechanisms of neural network predictions are complicated, and a lack of understanding
of causality between inputs and predicted values can limit the utility of these models. Olden et al.
(2004) explored techniques to assess the importance of variables within a neural network, but
the considered techniques were all computationally expensive. SHapley Additive exPlanations
(Lundberg and Lee 2017) and Local Interpretable Model-Agnostic Explanations (Ribeiro et al.
2016), are popular and efficient methods for determining feature importance. LIME has also been
extended to aggregate several realizations to explain the behavior of the whole model, which can
be used to detect general significant features in a neural network (Ribeiro et al. 2016).
Relative to prior work, this study extends surrogate models to substitute a more complex
traffic computation in large road networks. Furthermore, recent algorithms for determining feature
importance are adopted and extended to identify bridges that contribute most to network disruption
risk.

METHODOLOGICAL OVERVIEW
The study methodology consists of four main steps: (1) seismic risk analysis and data generation,
(2) traffic performance estimation using neural networks, (3) development of a modified LIME
algorithm (LIME-TI) to identify bridges for retrofit, and (4) evaluation of the performance of the
proposed retrofitting strategy. This process is presented in Figure 1.
The first step involves data generation through seismic risk assessment of the road network.
The data consist of realizations of traffic performance associated with damaged versions of the
road network due to seismic scenarios, generated through four sub-steps: (1) selecting seismic
scenarios; (2) obtaining realizations of the damaged road network; (3) using a traffic model, traffic
demand is assigned over the network; and (4) computing a traffic performance metric.
The second step is to build a neural network surrogate model to compute traffic performance.
The neural network architecture is defined by selecting hyperparameter values and evaluating neural
network prediction accuracy. This architecture definition is performed iteratively until accuracy
is maximized. Once hyperparameters have been selected, a calibrated neural network has been
defined.
The third step uses the neural network and a modified LIME variable importance algorithm
(LIME-TI) to predict which bridges have the greatest contribution to the system’s performance.
This ranking is used to identify bridges that could be retrofit to reduce seismic risk.
Finally, the fourth step evaluates the road network’s performance when subject to the retrofitting
actions from the third step.

3
 Traffic performance
Seismic risk analysis assessment using
and data generation neural networks
Select seismic scenarios Define Input & Output

Select hyperparameters
Sample damage maps
and sampling protocol

Assign traffic demand Define calibrated neural

network

Compute traffic
performance metric Modified LIME as a
retrofitting strategy
Sample realizations
using neural network

Compute LIME individual

explanations

Rank bridges using

Neural networks to
LIME-TI
predict seismic risk
Estimate seismic risk Evaluate retrofitting
parameters strategy

Fig. 1. Diagram of the process used in this study. Each part of the process matches a section in
this paper.

To demonstrate the process proposed in this study, the road network of the San Francisco Bay
Area is used as a testbed. This network is modeled by a directed graph containing 32,858 edges,
11,921 nodes, and 1743 bridges, as shown in Figure 2. This model was developed by Miller (2014),
and it was modified for this study by updating and correcting properties of some links.

SEISMIC RISK ANALYSIS AND DATA GENERATION

The first step of the methodology, as shown in Figure 1, is to generate training data by performing
a seismic risk analysis of the road network. Seismic risk assessment starts by selecting hazard
consistent seismic scenarios, which means sampling the scenarios considering their probability of

4
Fig. 2. Road network of the San Francisco
5 Bay Area used in this study.
 Δ𝑡𝑝 = 10.4%

1- Seismic Scenario 2- Damage Map 3- Traffic Assignment 4- Traffic Performance Metric

Fig. 3. Four steps of seismic risk assessment of road systems for a given seismic scenario.

occurrence according to the geological and seismological characteristics of the region. For each
scenario, the computation of network performance follows the four sub-steps illustrated in Figure
3.
The location and magnitude of the seismic scenario, along with local site information, are used
to feed a ground motion model (GMM) and spatial correlation model to obtain values of ground
motion intensity measure (IM) at the locations of all bridges are applied. For this study (i.e., the San
Francisco Bay Area road network), the IM used is the spectral acceleration at a period of 1 s. The
seismic scenarios are obtained using the OpenSHA Event Set Calculator (Field et al. 2003) with
the seismic source model developed by Field et al. (2009), near-surface shear-wave velocity (𝑉𝑠30 )
obtained by the method from Wald and Allen (2007), the Boore and Atkinson (2008) GMM, and
the model of Jayaram and Baker (2009) for spatial correlation in the IM values. These parameters
and models are consistent with the study of Miller (2014), upon which this work builds.
For a given sampled value of intensity measure, the following fragility function is used to
compute the probability of damage for each bridge in the system.


ln 𝑦/(𝛼𝜆 𝑘,𝑖 )
𝑃(𝐷𝑆𝑖 ≥ 𝑑𝑠 𝑘 |𝑌𝑖 = 𝑦) = Φ (1)
𝛽 𝑘,𝑖
where:
𝐷𝑆𝑖 = Damage state of component 𝑖.
𝑑𝑠 𝑘 = Damage state 𝑘.
𝑌𝑖 = Ground motion intensity measure value at the location of component 𝑖.
𝑃(𝐷𝑆𝑖 ≥ 𝑑𝑠 𝑘 |𝑌𝑖 = 𝑦) = Probability of component 𝑖’s damage state being 𝑑𝑠 𝑘 or greater, given
𝑌𝑖 = 𝑦.
Φ = Standard normal cumulative distribution function.
𝛼 = Factor that indicates retrofitting action: 𝛼 = 1 for no retrofit and 𝛼 = 1.2 for retrofit.
𝜆 𝑘,𝑖 = Median of 𝑌𝑖 causing damage state 𝑘 or greater.
𝛽 𝑘,𝑖 = Standard deviation of 𝑙𝑛(𝑌𝑖 ) causing damage state 𝑘 or greater.

6
 Damage states of each bridge are sampled with the above probabilities to obtain a damage
map. For this study, the relation between damage state of bridges and their traffic functionality
is consistent with Werner et al. (2006), who assigned functionality to the damage states proposed
by HAZUS (MR 2003). Specifically, no traffic is allowed on bridges in an extensive or complete
damage state, and bridges with no, slight, or moderate damage are still functional. For each seismic
scenario, several realizations of damage maps are obtained. For the San Francisco Bay Area, the
fragility curve parameters (i.e., 𝜆 𝑘,𝑖 and 𝛽 𝑘,𝑖 ) were provided by Caltrans (Miller 2014).
On the damaged version of the network, traffic demand is assigned using an Iterative Traffic
Assignment Algorithm (Chen and Alfa 1991). This algorithm takes fractions of the total demand
for each pair of origins and destinations and assigns it to the current shortest path between them.
Then, after the traffic has been assigned, the flow on each road of the network is updated, which
modifies the travel time on the road segment according to Equation (2), (Beckmann et al. 1956).
 4!
𝑞𝑎
𝑡 𝑎 = 𝑡 𝑓 1 + 0.15 (2)
𝑐𝑓
Where:
𝑡 𝑎 = Travel time of the road after flow assignment.
𝑡 𝑓 = Free flow travel time of the road.
𝑞 𝑎 = Current flow on the road.
𝑐 𝑓 = Capacity of the road.
This process is performed for each iteration and each pair of origins and destinations. Once all
users have been assigned a path, the aggregated travel time of the network is computed by adding all
travel times. The iterative traffic assignment is computationally expensive and imposes challenges
in evaluating seismic risk on these systems, especially when this process is performed to assess the
effectiveness of each retrofitting strategy.
Bridge damage may also disconnect some areas and prevent some trips from being completed.
To account for these lost trips, the following traffic performance metric (𝑡 𝑝) is used.

𝑡 𝑝 𝑘 = 𝑡𝑡 𝑘 + 𝛾𝑛𝑙𝑡 𝐾 (3)
Where:
𝑡 𝑝 𝑘 = Traffic performance metric for damage map 𝑘. This metric has units of time.
𝑡𝑡 𝑘 = Aggregated travel time for users for damage map 𝑘.
𝑛𝑙𝑡 𝑘 = Number of lost trips due to lack of connectivity for damage map 𝑘.
𝛾 = Penalty factor for lost trips. This study considers 𝛾 = 4 hours.

The percent change in the traffic performance metric Δ𝑡 𝑝 for damage map 𝑘, with respect to the
undamaged condition 𝑈𝐷, is defined as:
𝑡 𝑝 𝑘 − 𝑡 𝑝𝑈𝐷
Δ𝑡 𝑝 𝑘 = × 100 (4)
𝑡 𝑝𝑈𝐷

Where 𝑡 𝑝 𝑘 is the traffic performance metric shown in Equation (3) for damage map 𝑘, and 𝑡 𝑝𝑈𝐷
is the traffic performance metric in the undamaged condition which is equivalent to the aggregated
travel time for all users.

7
 The above process of computing network performance is performed for many scenarios. The
results are used as training data in the following section. They can also be combined with the
occurrence rates of the scenarios to compute Loss Curves, which measure the annual rate of
exceedance of a given network performance metric value:
𝑛
∑︁
𝜆 Δ𝑡 𝑝≥Δ𝑡 𝑝 0 = 𝑤 𝑘 I(Δ𝑡 𝑝 𝑘 ≥ Δ𝑡 𝑝0) (5)
𝑘=1
Where:
Δ𝑡 𝑝0 = Some level of increase in traffic performance metric.
𝜆Δ𝑡 𝑝≥Δ𝑡 𝑝 0 = Annual rate of events with Δ𝑡 𝑝 ≥ Δ𝑡 𝑝0.
𝑤 𝑘 = Annual rate of occurrence of damage map 𝑘 (based on the occurrence rate of the associated
seismic scenario).
𝑛 = Number of damage maps considered.
I( ) = an indicator function equal to 1 if the argument is true and 0 otherwise.

Another indicator of network performance is the expected annual increase in the traffic
performance metric:
𝑛
∑︁
𝐸 [Δ𝑡 𝑝] = 𝑤 𝑘 Δ𝑡 𝑝 𝑘 (6)
𝑘=1

Where:
Δ𝑡 𝑝 𝑘 = Level of increase in traffic performance metric for damage map 𝑘.
𝑤 𝑘 = Annual rate of occurrence of damage map 𝑘 (based on the occurrence rate of the associated
seismic scenario).
𝐸 [Δ𝑡 𝑝] = Expected annual increase in traffic performance metric.

This aggregated measure allows comparing the effects of different retrofitting strategies. In this
study, loss curves and expected annual increases in travel time are called "seismic risk parameters"
and are used to evaluate the accuracy of the neural network.
TRAFFIC PERFORMANCE ASSESSMENT USING NEURAL NETWORKS
The second step of Figure 1 is to develop a neural network as a surrogate model for the traffic
model. This model will take bridge damage states as input and output the traffic performance metric
of Equation 3, enabling a fast calculation of traffic parameters in a damaged road network. That is,
the neural network replaces Step 3 of the seismic risk assessment process (see Figure 3), which is
the most computationally expensive step.
Note that a neural network could be formulated to develop an End-to-End surrogate (Nabian
and Meidani 2018b). However, a neural network that skips more steps of the process will increase
the uncertainty in the predicted results and decrease the model’s accuracy. In this case, generating
realizations of damage given values of intensity measures is not computationally expensive, and
hence there is no benefit to skip Step 2 while decreasing the reliability of the model. This study
thus develops a neural network that only replaces the step of the seismic risk analysis with high
computational cost.

8
 To begin explaining the implementation of the neural network used to predict traffic perfor-
mance, some general methodological considerations for neural network calibration will be intro-
duced. Then, the application of each consideration to the San Francisco Bay Area testbed is
shown.
Methodology for neural network calibration
Several aspects need to be considered to train, validate and test a neural network. The most
significant for this study are (1) definition of inputs and outputs, (2) selection of sampling protocols
for training data, (3) selection of model hyperparameters, and (4) proposal of a evaluation method
that is scientifically and practically insightful. The following sections elaborate on these issues.
Definition of inputs and outputs
The first step to develop a neural network is to define the inputs and outputs of the model.
One aspect that should be considered when defining these variables is that the computational time
required to compute the outputs of the model should be reduced significantly. In addition, to
improve network accuracy, randomness in the inputs should be avoided.
Selecting sampling protocols for training data
As part of the calibration process, the sampling protocol that will be used to generate training
data must be defined. The generated data should represent conditions that the neural network is
expected to encounter during its utilization. Some challenges that should be accounted for during
the sampling protocol definition are the selection of a training data distribution that minimizes the
neural network’s bias on test data. In disaster engineering, the sampling process should consider
that highly disruptive events are of great interest but occur much less frequently than non-disruptive
events.
Hyperparameters selection
Once the input and outputs of the model have been selected, the parameters that define the
neural network, or hyperparameters, need to be determined. These include the number of layers
in the neural network, the optimization algorithm, the learning rate of the optimization algorithm,
and the number of neurons per layer. In addition, the connectivity between neurons across different
layers of the neural network must be determined.
One common approach to determine hyperparameter values is to perform a grid search in
which all possible combinations of hyperparameters are evaluated in terms of the accuracy of the
neural network on test data. The hyperparameter combination that minimizes the error can then be
selected. For each sampling protocol, a hyperparameter selection should be conducted to minimize
accuracy error on test data.
Evaluation of the calibrated neural network
The final step of calibrating a neural network is its accuracy evaluation. Through this process,
the neural network’ suitability as a surrogate model is tested. The evaluation is traditionally
performed by analyzing whether its accuracy is within an acceptable range for the specific study.
Example metrics that can be used to evaluate the performance of regression neural networks are
logarithmic loss, mean squared error, mean absolute error, and coefficient of determination 𝑅 2 .
The developer of a neural network can additionally propose other evaluation metrics focused on
the specific study that is being conducted. For example, instead of aggregate metrics, the developer
can propose metrics that focus on cases that are critical to the model. For instance, in structural

9
engineering, it could be essential to focus on the capabilities of a model to predict structural collapse
more accurately than another kind of structural response. In this study, these other metrics are
named "engineering parameters."
Neural network calibration for the San Francisco Bay Area road network
The following subsections present how the previous calibration considerations for neural net-
works were implemented in a traffic model for the San Francisco Bay Area road network.
Definition of inputs and outputs
The neural network proposed in this study takes damage states of bridges (input) to compute
a metric of overall network performance (output). The damage state of each bridge is represented
by a binary variable, which is 1 when the bridge is extensively damaged or worse, and is 0 when
not. The output of the neural network is the change in traffic performance metric as presented in
Equation (3).
In selecting adequate features, topological features of the network graph (such as edge degree,
centrality, or graph diameters) were also analyzed as potential inputs. However, including these
parameters did not significantly improve the performance of the trained network. Moreover, in the
case of some centrality measures, the time involved in calculating these measures was even higher
than running the traffic model, which invalidated the purpose of the neural network itself. Hence,
only bridge damage states are used as inputs for the neural network.
Selecting sampling protocols for training data
This study considered three protocols to generate training data: (1) sampling hazard consistent
seismic scenarios; (2) over-sampling scenarios that generate severe consequences, named here as
Extreme Events; and (3) random sampling of scenarios. The first two sampling protocols align
more directly with the intended risk analysis goal, whereas the random sampling protocol is used
as a benchmark to compare the performance of the other two protocols. The second protocol is
considered because preliminary results indicated that the first protocol did not produce enough
highly disruptive events, making the prediction of extreme events less accurate. In all three
protocols, simulations with no bridge damage were neglected since they do not provide helpful
information for training the neural network.
An extreme event is characterized here as an event that generates an increase in traffic perfor-
mance metric bigger than 50% of the most significant disruption in the hazard consistent sampling.
Given that this categorization is somewhat arbitrary, different values of the extreme event thresholds
were explored. The final value of 50% was selected because it provided a balance between gen-
erating enough data from hazard consistent seismic scenarios while still corresponding to highly
disruptive events. With this threshold, approximately 6% of the damage maps obtained from
hazard-consistent sampling are defined as extreme events. In other applications, instead of using a
metric relative to the maximum observed disruption, this categorization can be defined considering
the decision-maker’s acceptable performance threshold and should produce comparable results.
To study the effect of over-sampling Extreme Events when training the neural network, several
training sets of the same size with different proportions of extreme events were considered by
drawing samples of data from the extreme and non-extreme categories with varying proportions.
The original pool from which these samples were drawn was larger than the final training data to
allow a sufficient number of data to sample from. Specifically, a set that was one order of magnitude
bigger than the biggest amount used to train the hazard consistent model was used.

10
 0.95

0.90

0.85
R2

0.80

0.75

0.70
All events in test data
Extreme events in test data
0.0 0.2 0.4 0.6 0.8 1.0
Fraction of training data corresponding to extreme events
Fig. 4. Coefficients of determination computed considering all realizations in the test dataset and
considering realization consistent with the definition of extreme events.

The varying sets of training data were then used to train the neural network, and its performance
was measured using the coefficient of determination (𝑅 2 ) on test data sets. Separate test data
sets were compiled that consisted of hazard-consistent events or that consisted of only extreme
events. The latter test set allows to directly evaluate the neural network’s performance for extreme
events–the events of great interest for seismic risk management activities. Performance for varying
sampling protocols and test data sets are shown in Figure 4. The accuracy in predicting the All
events test data is relatively stable for training sets with ≤ 90% extreme events but decreases for
larger percentages. As expected, the accuracy in predicting the extreme events test data increases
as the proportion of extreme events increases in the training data, peaking for training data with
≥ 80% extreme events. Thus, the extreme event protocol used for the remainder of this work
consists of a training set with a proportion of extreme events of 80%, as in this case, the resulting
neural network performs well for both extreme events and all events. Note that the 80% extreme
events is a much larger proportion than the 6% of extreme events produced by the hazard consistent
protocol. The improved performance from this over-sampling (as seen in Figure 4) was also seen
to significantly improve the performance of the neural network in later analysis stages.

11
 Hidden Layers

DS1

DS2

Change in
Features

traffic
performance
.
metric Δtp
.
.
.

DSN

Fig. 5. Illustration of the neural network architecture. The neural network goes from a binary
indicator for each bridge, through 10 hidden layers with 150 neurons each, to obtain the change in
traffic performance metric.

Hyperparameters calibration
This study performed a hyperparameters calibration of the number of neurons per layer, the
number of layers in the neural network, the learning rate of the algorithm, and the L2 penalty factor.
These hyperparameters define the architecture of the neural network, which is illustrated in Figure
5. In addition, the calibrated neural network was considered to be fully connected. The neural
network was trained using the ADAM (Adaptive Moment Estimation) (Kingma and Ba 2014). This
algorithm was chosen because it converged faster than others, such as stochastic gradient descent
(SGD). The rectified linear unit function (ReLU) was used for all hidden layers as the activation
function. The values of hyperparameters that minimized test error are summarized in Table 1.
Despite performing the hyperparameters calibration for each sampling protocol, there were no
significant differences between them; therefore, the same hyperparameters were used throughout
different sampling protocols.
Evaluation of the calibrated neural network
Using the hyperparameters shown in Table 1, the neural network for each protocol is evaluated
using statistical evaluation for neural networks. The first analysis conducted is a measure of model
accuracy through the use of 𝑅 2 . The results of the different regression models are shown in Table
2. Extreme events sampling and hazard-consistent event sampling have a value of 𝑅 2 of 0.967
and 0.958, respectively, on the test data, which has not been used to train nor validate the neural
network. To verify how good this value is, a linear regression and a ridge regression is performed
using the same training data, observing 𝑅 2 values of 0.842 and 0.864, respectively. To conduct
this regression, we use all the damage states of the bridges as variables in the same way as it is
done for the neural network. Thus, neural networks produce more accurate results than other more

12
 TABLE 1. Hyperparameters used to calibrate the neural network model

Hyperparameter Values
Optimization algorithm ADAM
Activation function ReLU
Number of layers 10
Number of neurons per layer 150
Learning rate 𝛼 0.001
Penalty Factor L2 0.001

TABLE 2. Coefficient of determination, as a metric of accuracy, for different regression models

and sampling protocols.

Regression Model Coefficient of determination 𝑅 2

Neural Network - Extreme Events 0.967
Neural Network - Hazard Consistent 0.958
Neural Network - Random Sampling 0.888
Linear Regression 0.842
Ridge Regression 0.864

straightforward regression techniques.

As mentioned before, the motivation to include a sampling protocol focused on extreme events
was to accurately represent events that can be particularly disruptive and of interest to decision-
makers. Taking this into account, to show the difference between the predicted values of the
neural network that uses hazard-consistent sampling and the one that uses extreme events sampling,
Figure 6 compares the Δ𝑡 𝑝 values predicted by the neural network and by the traffic model for
both sampling protocols on the test set, showing all test data in grey and events that are extreme
in blue. As expected, for extreme events, the prediction performed by extreme events sampling
is more accurate than the one performed by hazard-consistent sampling. On the other hand, the
prediction of extreme events sampling for non-extreme events is not as accurate as the prediction
given by the hazard consistent sampling protocol. Considering that the events that are not that
extreme have a smaller role in the risk of road systems, the observed difference is not significant.
Possible overfitting of extreme events in Figure 6 was checked by taking several sets of extreme
events not used on the training data, and the value of 𝑅 2 remained stable and lower than the one
observed on the training data. Whether or not the increase in accuracy over extreme events was
a result of just including more extreme events into the training set was also explored, concluding
that it was not the result of an increase of training data but rather an increase in the proportion
of extreme events with respect to the total. This dependence on the proportion is consistent with
assigning bigger weights to extreme events while minimizing the training error in the calibration

13
Fig. 6. Scatter plots showing a comparison between predicted values by the neural network, and
data obtained using the traffic model. Results for hazard consistent protocol are shown on the left
and for extreme events on the right.

of the neural network. The sudden drop observed in Figure 6 is a result of the definition of extreme
events and their consequences on the sampling protocol. If other thresholds to characterize extreme
events were defined, the sudden drop would be observed in other values of network performance;
on the other hand, if a probability distribution was assigned to the sample of extreme events that
were not binary and defined by a threshold, then the drop could be minimized. For the effects of
this study, the sudden drop does not affect the purpose of the neural network calibration, which is
to allow seismic risk assessment. The previous statements regarding how to control the location
and magnitude of the drop observed on Figure 6 depend on specific implementations of the reader.
To further illustrate the benefit of extreme events sampling, Figure 7 shows the residuals as a
function of Δ𝑡 𝑝 using each model. It is observed that the hazard-consistent protocol has a more
considerable bias and that the difference in bias is more significant when predicting extreme events.
Note that random sampling is the protocol that leads to the smallest bias for results greater than
75 [%UD]; however, its accuracy is significantly worse than the other models when observing
Figure 8. Given the test data structure and how 𝑅 2 is computed, there is a trade-off between bias
and accuracy. Test data, which represents how modelers will evaluate the network, has seismic
events that are hazard consistent, which means an underlying distribution of the data focused on
minor events. This distribution is different from the implicit uniform distribution used to compute
𝑅 2 since hazard consistency adds more weight to events with less disruption, which happen more
frequently. The trade-off between 𝑅 2 and the bias comes from this underlying distribution since
higher 𝑅 2 values will minimize the error for smaller events and at the cost of increasing the error for
disruptive events. Different training data distributions were considered, but none could minimize
the bias while obtaining a high value of 𝑅 2 .
Another experiment conducted to verify the network’s performance was to explore the accuracy
of the neural network as a function of the number of realizations included in the training data.
Both the neural network trained with extreme events sampling and hazard-consistent sampling are

14
 60
Random
Mean error in bin measured as Δtp

Extreme events
Δ0 Ha ard consistent

20

−20

−Δ0

−60
25 50 75 100 125 150 175
Change in traffic performance metric Δtp
Fig. 7. Mean residuals for extreme events computed using a neural network trained with each of
the three sampling protocols. Optimal zero bias shown with red dashed line.

significantly more accurate than the neural network trained using random sampling. Moreover,
Figure 8 shows that after approximately 10000 realizations, the neural network’s performance
remains stable. This information suggests that although it may be computationally expensive to
train the network, it is feasible, and the benefits of using the neural network surpass this cost. For
reference, for a hazard consistent seismic assessment of the San Francisco Bay Area, 1992 scenarios
are used as provided by implementing UCERF2 according to Field et al. (2009). Therefore, training
this neural network is equivalent to performing five analyses, which is not a considerable number
taking into account that every time a single retrofitting strategy is evaluated, all those 1992 scenarios
need to be considered.
As a conclusion of the comparison between the neural networks trained with the three sampling
protocols, the neural network that uses extreme events can predict Δ𝑡 𝑝 as well as the neural network
that uses hazard-consistent sampling, without inducing a bias on the extreme events, which are
particularly important for decision-making in risk management. Given this, the following sections
only present results of the neural network trained using extreme events sampling.
MODIFIED LIME ALGORITHM AS A RETROFITTING STRATEGY

15
 0.9
Value of R 2 on test data

0.8

0.7

Extreme events
0.6
Hazard consistent
Random
103 104 105 106
Number of realizations in training data
Fig. 8. Evolution of neural network accuracy as a function of the number of realizations included in
the training data. Results are shown for random, extreme-events, and hazard-consistent sampling.

In addition to predicting the performance of the damaged road network, the second goal of
this study is to use the neural network to identify the bridges whose damage contributes most to
network disruption. Given that the input variables are the bridge damage states, the influential
bridges can potentially be identified by evaluating the most important input variables in predicting
traffic performance. By ranking the importance of the bridges in the model, a new retrofitting
strategy can be proposed. This calculation is the third step of the Figure 1 framework.
LIME as a variable importance algorithm
To establish which variables within a model have a more prominent role in predicting its output,
it is necessary to use a variable importance algorithm. In that regard, this study uses a Local
Interpretative Model-Agnostic Explanation (LIME) to efficiently determine which variables had a
more significant role in predicting the output variable (Ribeiro et al. 2016). The LIME algorithm
develops local linear regressions around a singular realization of interest using the input variables
of the neural network as independent variables and the values predicted by the neural network
around the realization as dependent variables. This LIME algorithm was initially meant to explain
individual predictions by the neural network; however, it can be used in several realizations to

16
extrapolate which variables contribute more frequently to the neural network model as a whole.
This process, named "submodularity pick" in Ribeiro et al. (2016), consists of three steps: (1) select
a set of individuals realizations to be explained by the LIME algorithm, (2) compute coefficients
𝑊 𝑗 𝑘 for each feature on each realization, and (3) aggregate for each feature the coefficients obtained
for every single realization. As a result of this algorithm, each independent variable, which
corresponds to a bridge, has a weight 𝐼 𝑗 that describes how much it contributes to the whole neural
network. In the original formulation of LIME, the coefficients of the features 𝑘 for each realization
𝑗 are the coefficients of a local linear regression 𝑊 𝑗 𝑘 . Then, following step 3, the global coefficient
of a feature 𝐼 𝑗 is defined by
v
u
t 𝑑
∑︁
𝐼𝑗 = |𝑊 𝑗 𝑘 | (7)
𝑘=1

Where
𝐼 𝑗 = Final weight of variable 𝑗.
𝑊 𝑗 𝑘 = Weight of variable 𝑗 in realization 𝑘.
𝑑 = Number of realizations considered in the subsampling.

Bridges are ranked according to their 𝐼 𝑗 values, which represent their importance in estimating
traffic disruption.
LIME accounting for Traffic Impacts (LIME-TI)
The original submodularity pick calculation above was revised to be more suitable for this
application. The modified approach, termed LIME with Traffic Impacts, incorporates the bridge’s
contribution to traffic performance and the probabilistic nature of seismic events. To account
for these two factors, Steps 2 and 3 of LIME were changed. To incorporate bridge importance
in traffic performance, Step 2 is modified by computing the coefficients 𝐶𝑖 𝑗 of each bridge 𝑗
on each realization 𝑖 as the marginal increase in traffic performance once the bridge has been
repaired according to the order suggested by the local explanation of LIME for the realization
𝑖, as shown in Equation (8). For instance, if a realization has two bridges A and B, and the
explanation of the realization from LIME establishes that bridge A has higher importance than B,
then the Traffic Impacts (TI) coefficient for bridge A, 𝐶𝑖 𝐴 , will be the improvement in network
performance metric Δ𝑡 𝑝 when only bridge A is repaired. Regarding bridge B, its TI-coefficient 𝐶𝑖𝐵
will be the improvement on the traffic performance metric Δ𝑡 𝑝 after repairing B, given that A was
already repaired. This conditioning on having repaired more important bridges before is why the
TI-coefficient is defined as the marginal improvement on the traffic performance metric.

𝐶 𝑗 𝑘 = 𝑡 𝑝 𝑘 − 𝑡 𝑝 𝑏≥ 𝑗,𝑘 (8)
Where
𝐶 𝑗 𝑘 = TI-coefficient of bridge 𝑗 in realization/damage map 𝑘.
𝑡 𝑝 𝑘 = Traffic performance metric computed for damage map 𝑘.
𝑡 𝑝 𝑏≥ 𝑗,𝑘 = Traffic performance metric computed when all bridges with a higher LIME ranking than
bridge 𝑗, and bridge 𝑗 itself have been repaired for damage map 𝑘.

17
 In addition to redefining LIME coefficients, this study also modifies how these coefficients
are aggregated in Step 3. In place of using Equation (7), which adds all realizations equally, the
realizations are now combined using their rates of occurrence 𝑤 𝑘 as shown in Equation (9). By
aggregating all realizations using 𝑤 𝑘 , each bridge contribution is normalized by their vulnerability
and the occurrence rate of seismic scenarios.
𝑛
∑︁
𝐼 ∗𝑗 = 𝑤 𝑘 |𝐶 𝑗 𝑘 | (9)
𝑘=1
Where
𝐼 ∗𝑗 = LIME-TI coefficient of bridge 𝑗.
𝐶 𝑗 𝑘 = TI-coefficient of bridge 𝑗 in realization/damage map 𝑘.
𝑤 𝑘 = Rate of occurrence of damage map 𝑘 as used in (5) and (6).

The result of this LIME-TI algorithm is a ranking of bridges that are important to predict the
traffic disruption, but it also accounts for their vulnerability and the contribution of other damaged
bridges.
Evaluation of retrofitting strategies on the San Francisco Bay Area road network
Once the ranking from LIME-TI is obtained, as shown in Figure 1, its effectiveness to reduce
road network disruption is compared to those of rankings obtained using three other potential
metrics of importance: structural vulnerability, one-at-a-time (OAT) sensitivity analysis, and traffic
capacity of the bridges. The vulnerability ranking is obtained by sorting the number of times the
bridge was damaged in all training data scenarios. The more times a bridge experiences damage,
the higher its vulnerability ranking. The OAT analysis uses the value of Δ𝑡 𝑝 generated by damaging
each bridge individually; then, bridges are ranked in descending order according to the Δ𝑡 𝑝 that
their damage causes. Finally, the traffic capacity ranking is based on the amount of traffic that the
bridge allows in free-flow conditions. Rankings from the original LIME formulation were also
checked, but results are not reported since they were less effective than LIME-TI.
The retrofitting strategies were evaluated by computing the expected annual change in road
performance metric 𝐸 [Δ𝑡 𝑝] for different numbers of bridges retrofitted. For a case of 𝑛 retrofits,
the 𝑛 highest-ranking bridges according to each metric were assumed to be retrofit. A retrofit bridge
has its fragility function in Equation 1 modified, and the entire risk assessment procedure is repeated
(with the retrofit bridge being damaged less often). 𝐸 [Δ𝑡 𝑝] is computed using Equation (6). The
performance evaluation considers many scenarios; the traffic performance metric is computed using
the neural network surrogate model to reduce computing time. The results of this evaluation are
presented in Figure 9, which shows that the ranking based on LIME-TI submodularity yields a
bigger improvement in traffic performance than ranking bridges by structural vulnerability, OAT
analysis, or traffic capacity.
Explanation on bridges selected by LIME-TI
The superiority of the proposed LIME-TI ranking is reasonable as it is the only ranking that
combines the effect of vulnerability and traffic importance. OAT analysis and traffic capacity only
quantify the role of the bridge in traffic performance, whereas structural vulnerability only accounts
for bridge damage probability without including the importance of the bridge to the system’s
performance.

18
 Centrality
5 Vulnerability
Expected Annual Travel Time Increase [%]

Neural Network Ranking

OAT
4

0
0 5 10 15 20 25 30 35
Percentage of Retrofitted Bridges

Fig. 9. Expected annual increase in traffic performance metric for different retrofitting strategies
as a function of the number of retrofitted bridges.

This study next explores the characteristics of the bridges that the LIME-TI algorithm ranks
highly. There was no strong correlation between the LIME-TI ranking and the traffic capacity
of the bridges or their structural vulnerability, separately. However, some correlations can be
observed with the bridges selected by LIME-TI when using combined traffic capacity and structural
vulnerability measures. This correlation exists because some bridges may have an important role
in the network, but they may not be very vulnerable and contribute significantly to the seismic risk.
On the other hand, the most vulnerable bridges in the network may not have an essential role in
traffic performance; therefore, retrofitting those bridges is not effective in minimizing the impacts of
earthquakes. The LIME-TI selection of important bridges comes from the magnitude of coefficients
𝐶 𝑗 𝑘 , which detects the vulnerability of bridges by the occurrence of bridge damage on the individual
realizations, combined with the occurrence rates used to aggregate different realizations in Equation
(9). Motivated by this correlation, rankings based on combinations of OAT and vulnerability were
also evaluated, but even combinations of these factors did not outperform the LIME-TI ranking.
Besides correlations, the features that make the bridges proposed by LIME-TI improve the
network’s performance better than other strategies were also explored. The first bridges selected by
LIME-TI have a more significant individual contribution to the expected annual increase of traffic

19
 All bridge
Expected traffic performance metric

6 Top 50 bridge from OAT

Top 50 bridge prioritized by LIMEΔTI
5
per bridge E[Δtp]

0
0 10 20 30 40 50
Traffic performance metric [Δtp] from OAT analy i
Fig. 10. Scatter plot of traffic performance from OAT analysis versus expected traffic performance
metric, for all bridges in the network. The top 50 bridges ranked by OAT are shown in blue and the
top 50 bridges ranked by LIME-TI are shown in orange.

performance metric of the traffic network. Figure 10 shows on the X-axis the change in traffic
performance when each bridge is damaged individually, which is defined in previous sections as
the OAT analysis. The Y-axis shows the expected change in traffic performance metric per bridge
as defined in Equation (10).
∑︁
𝐸 [Δ𝑡 𝑝] 𝑗 = 𝑤𝑘 𝑥 𝑗 𝑘𝑇 𝑃 𝑗 𝑘 (10)
𝑘
Where:
𝐸 [Δ𝑡 𝑝] = Expected change in traffic performance metric associated to bridge 𝑗 .
𝑤 𝑘 = Rate of occurrence of damage map 𝑘.
𝑥 𝑗 𝑘 = Indicator of damage to bridge 𝑗, in damage map 𝑘. = 1 if the bridge is damaged, 0 otherwise.
𝑇 𝑃 𝑗 𝑘 = Contribution of bridge 𝑗 to traffic performance metric change in scenario 𝑘, defined in
Equation (11).

20
 Δ𝑡 𝑝 𝑘
𝑇𝑃𝑗𝑘 = (11)
𝑛𝑏 𝑘
Where:
𝑇 𝑃 𝑗 𝑘 = Contribution of bridge 𝑗 to traffic performance metric change in scenario 𝑘.
Δ𝑡 𝑝𝑘𝑖 = Change in traffic performance metric for scenario 𝑘
𝑛𝑏 𝑘 = Number of bridges damaged in scenario 𝑘

Figure 10 shows that the bridges selected by LIME-TI may not individually cause substantial
disruption to the network when they experience damage. However, they generate significant
traffic disruption in the seismic scenarios considered on the training data, resulting from their
vulnerability and importance to the network.

Additionally, the locations of the bridges selected by LIME-TI within the road network are
examined. It is observed that LIME-TI bridges often correspond to "leading bridges." A leading
bridge is a bridge that does not generate a significant change in 𝑡 𝑝 when it experiences damage, but
that leads or connects to a bridge that, when damaged, does generate a considerable change in 𝑡 𝑝.
An OAT analysis did not identify these bridges since they are not the most important ones, and their
ranking in OAT is lower than the one assigned by LIME-TI, which accounts for their vulnerability.
Examples of leading bridges are shown in Figure 11, in which some of the top bridges selected
by LIME-TI are shown with blue circles. These bridges are on highways that lead to two critical
bridges in the San Francisco Bay Area, the Golden Gate Bridge and the Bay Bridge, which are not
particularly vulnerable in our model.

NEURAL NETWORKS TO PREDICT SEISMIC RISK

The neural network of this study is a tool to inform engineering decisions. Thus, besides the
evaluation of a neural network based on prediction errors, the model should predict seismic risk
results used by policymakers. In particular, this work studies the model’s accuracy at predicting
loss curves and expected annual loss for the system.
The first seismic risk metric that the neural network should estimate accurately is a loss curve,
as shown by Equation (5). Thus, the loss curves obtained using the traffic model described in
the previous section are compared to the curves computed using the neural network. Figure 12
shows the loss curves for the case without retrofitting and for the case where 300 bridges are
retrofitted based on their LIME-TI ranking. In both cases, the neural network can reproduce the
curve generated by running the traffic model.
Figure 13 presents the expected annual increase of traffic performance computed by using the
traffic model and by using the neural network model for different numbers of retrofitted bridges.
The sets of bridges to retrofit correspond to those from the LIME-TI ranking, which coincide with
those used in Figure 9, and the associated expected annual loss is computed using Equation (6).
The prediction obtained from the neural network has almost no bias and has low error relative to
the real value for all numbers of retrofitted bridges.
These results are significant because these risk metrics incur a high computational cost if
obtained using the traffic model. Thus, the fact that the neural network is highly accurate allows
reliable estimates of the effect of retrofitting bridges in a matter of seconds.

21
 Bridge not selected by LIME-TI
Bridge selected by LIME-TI

Fig. 11. A map of a portion of the testbed, illustrating "leading bridges." Bridges among the top 50
selected by LIME-TI are shown with blue dots. Other bridges are shown with black dots. Bridges
that could be leading bridges but that are not vulnerable enough to be highly ranked are marked in
a green box.

DISCUSSION
The results obtained in this study prove that neural networks can be extended to perform a seismic
risk assessment of distributed systems in a rapid and accurate way. In terms of time savings, the
neural network can perform analysis approximately 100,000 times faster than the original traffic
model. This significant decrease of computational time did not impact the model’s accuracy
significantly, with the coefficient of determination of the model being 𝑅 2 = 0.97. Moreover,
besides this statistical evaluation of the neural network, it was also observed that the surrogate
model could predict practical parameters for decision-making. In particular, the use of the model
allowed the fast computation of loss curves and expected annual losses.
Training neural networks to predict the performance of distributed systems introduces an alter-
nate way to efficiently assess risk of critical infrastructure. Prior to this study, scenario reduction
techniques, such as those proposed by Miller (2014) and (Han and Davidson 2012), have been used
as methods to optimize computational resources without compromising the accuracy of the risk
assessment. With the introduction of surrogate models such as the ones proposed in this work, the

22
Annual rate of e ceedance

10−2

10−3

Traffic model - No retrofitting

Neural network - No retrofitting
Traffic model - 300 bridges
Neural network - 300 bridges
0 250 500 750 1000 1250 1500 1750 2000
Change in traffic performance metric Δtp
Fig. 12. Comparison of loss curves obtained using the neural network and the traffic model.

time required to analyze the consequences of each scenario is minimal, therefore limiting the need
to reduce the number of scenarios included in the risk assessment.
From a practical point of view, these rapid and accurate surrogate models can be instrumental
for decision-makers managing complex systems. The ability of the model to predict metrics such
as expected annual loss or loss curves in a short time can facilitate the exploration of management
strategies without requiring access to high-performance computing resources.

CONCLUSIONS
This study proposes a neural network surrogate model that accurately and rapidly estimates
traffic disruption due to earthquakes. The neural network uses damage states of all bridges to
predict a traffic performance metric and achieves a value of 𝑅 2 = 0.97 when compared with the
traffic model in the testbed study, but with greatly reduced computational time. In addition, the
neural network performs well based on risk measures such as loss curves and expected annual loss,
which is relevant for evaluating the impact of retrofitting strategies.
In addition, the study proposes a new retrofitting strategy based on ranking bridges according
to their role in the disruption to the road network. To quantify the role of the bridges, this work
uses a modified version of a Local Interpretable Model-Agnostic Explanation (LIME), which is a

23
 45
E[Δtp]ΔusingΔneuralΔnetworkΔmodel

40

35

30

25

20
20 25 30 35 40 45
E[Δtp]ΔusingΔtrafficΔmodel
Fig. 13. Comparison of expected annual loss computed using the neural network and using the
traffic model. Each point represents a different number of retrofitted bridges consistent with the sets
of bridges used in Figure 9. Points with higher 𝐸 [Δ𝑡 𝑝] values are associated with fewer retrofitted
bridges. The diagonal line indicates cases where the neural network results perfectly predict the
traffic model result.

variable importance algorithm applicable to neural networks. In the modified version (LIME-TI),
a new submodularity pick that accounts for the vulnerability of bridges and their importance in
the performance of the road network is proposed. The proposed retrofitting strategy yields smaller
increases in expected traffic disruption than other intuitive retrofitting strategies, such as ranking
bridges by vulnerability or the volume of traffic they carry.
It is observed that some of the bridges selected by LIME-TI have a significant individual
contribution to the seismic risk of the network. Moreover, some of these bridges are termed a
leading bridge, a bridge whose sole impairment does not generate a significant disruption to the
network but leads to a bridge with a significant role in the road network.
To construct the neural network, three different sampling protocols to generate training data
are explored, and this study concludes that it is essential to incorporate enough highly disruptive
events in the training data to predict these infrequent events accurately. This predictive capability

24
is also important for the seismic management of road systems because decision-makers may rely
on scenario-based approaches to explore the impact of different retrofitting strategies, which are
usually selected to be highly disruptive. The authors also conclude that random sampling is less
accurate than using hazard consistent or extreme events sampling.
The surrogate model requires a significant number of network performance simulations for
training. However, the simulations are independent, and hence they can be performed in parallel.
In addition, the objective of the neural network is to serve as a tool to evaluate retrofitting policies.
As such, it still massively reduces the overall computational cost of the results shown above because
risk metrics were computed for several retrofit permutations. Each retrofit case would take several
hours or days of computational time to evaluate, whereas the trained neural network only takes
fractions of a second. In terms of policy, the network could be trained by a research institution or
consultant with access to high-performance computing and then transferred to decision-makers for
use. The number of damage realizations required to train the network above was equivalent to that
from evaluating the effectiveness of five retrofitting strategies. Thus, for any substantial number
of retrofit strategy evaluations, the surrogate model will greatly reduce the overall computational
cost. For instance, to obtain Figure 13, it takes 112.68 hours to run the original traffic model on
a 2016 MacBook with a 2.6 GHz Quad-Core Intel Core i7 processor. The neural network takes
10.89 seconds to generate the same data on the same computer.
This study indicates several considerations for future applications on other systems that should
allow the method to be generally valuable. First, properly trained neural networks can predict
flows in infrastructure networks and reflect complex relationships between component damage and
performance. Second, it is likely necessary to over-sample high-damage realizations of the network
during training to avoid over-fitting small damage cases and failing to effectively predict the high-
disruption cases of interest in risk management problems. Finally, a trained surrogate model offers
tremendous utility in allowing decision-makers to explore mitigation strategies such as component
retrofit by capturing complexities of the network performance at a feasible computational cost.

DATA AVAILABILITY STATEMENT

Some data, models, or code generated or used during the study are available in a repository
online in accordance with funder data retention policies, which can be found in: DOI 10.5281/
zenodo.5161259.

ACKNOWLEDGMENTS
The State of California supported this work through the Transportation System Research Pro-
gram of the Pacific Earthquake Engineering Research Center (PEER) and by the Shah Family Fund
Fellowship. Some of the computing for this project was performed on the Sherlock cluster. The
authors would like to thank Stanford University and the Stanford Research Computing Center for
providing computational resources and support that contributed to these research results. Any
opinions, findings, conclusions, or recommendations expressed in this material are those of the
authors and do not necessarily reflect those of the funding agency.

REFERENCES
Alemzadeh, S., Talebiyan, H., Talebi, S., Duenas-Osorio, L., and Mesbahi, M. (2020).
“Deep learning-based resource allocation for infrastructure resilience.” arXiv preprint
arXiv:2007.05880.

25
Basoz, N. and Kiremidjian, A. S. (1995). Prioritization of bridges for seismic retrofitting. National
Center for Earthquake Engineering Research.
Beckmann, M., McGuire, C. B., and Winsten, C. B. (1956). Studies in the Economics of Trans-
portation. Yale University Press, New Haven, CT.
Bommer, J., Scott, S., and Sarma, S. (2000). “Hazard-consistent earthquake scenarios.” Soil
Dynamics and Earthquake Engineering, 19(4), 219–231.
Bommer, J., Spence, R., Erdik, M., Tabuchi, S., Aydinoglu, N., Booth, E., Del Re, D., and Peterken,
O. (2002). “Development of an earthquake loss model for turkish catastrophe insurance.” Journal
of seismology, 6(3), 431–446.
Boore, D. M. and Atkinson, G. M. (2008). “Ground-motion prediction equations for the average
horizontal component of pga, pgv, and 5%-damped psa at spectral periods between 0.01 s and
10.0 s.” Earthquake Spectra, 24(1), 99–138.
Caltrans, S. (2019). “Caltrans seismic design criteria.” California Department of Transportation,
Sacramento, CA.
Chen, M. and Alfa, A. S. (1991). “A network design algorithm using a stochastic incremental traffic
assignment approach.” Transportation Science, 25(3), 215–224.
Elhag, T. M. and Wang, Y.-M. (2007). “Risk assessment for bridge maintenance projects: neural
networks versus regression techniques.” Journal of computing in civil engineering, 21(6), 402–
409.
Field, E. H., Dawson, T. E., Felzer, K. R., Frankel, A. D., Gupta, V., Jordan, T. H., Parsons, T.,
Petersen, M. D., Stein, R. S., Weldon, R., et al. (2009). “Uniform california earthquake rupture
forecast, version 2 (ucerf 2).” Bulletin of the Seismological Society of America, 99(4), 2053–2107.
Field, E. H., Jordan, T. H., and Cornell, C. A. (2003). “Opensha: A developing community-modeling
environment for seismic hazard analysis.” Seismological Research Letters, 74(4), 406–419.
Gomez, C. and Baker, J. W. (2019). “An optimization-based decision support framework for coupled
pre-and post-earthquake infrastructure risk management.” Structural Safety, 77, 1–9.
Gomez, C., Sanchez-Silva, M., Dueñas-Osorio, L., and Rosowsky, D. (2013). “Hierarchical in-
frastructure network representation methods for risk-based decision-making.” Structure and
Infrastructure Engineering, 9(3), 260–274.
Han, Y. and Davidson, R. A. (2012). “Probabilistic seismic hazard analysis for spatially distributed
infrastructure.” Earthquake Engineering & Structural Dynamics, 41(15), 2141–2158.
Jayaram, N. and Baker, J. W. (2009). “Correlation model for spatially distributed ground-motion
intensities.” Earthquake Engineering & Structural Dynamics, 38(15), 1687–1708.
Kingma, D. P. and Ba, J. (2014). “Adam: A method for stochastic optimization.” arXiv preprint
arXiv:1412.6980.
Lim, H.-W., Song, J., and Kurtz, N. (2015). “Seismic reliability assessment of lifeline networks
using clustering-based multi-scale approach.” Earthquake Engineering & Structural Dynamics,
44(3), 355–369.
Liu, C., Fan, Y., and Ordóñez, F. (2009). “A two-stage stochastic programming model for trans-
portation network protection.” Computers & Operations Research, 36(5), 1582–1590.
Lundberg, S. and Lee, S.-I. (2017). “A unified approach to interpreting model predictions.” arXiv
preprint arXiv:1705.07874.
Miller, M. (2014). “Seismic risk assessment of complex transportation networks.” Ph.D. thesis,
Stanford University, Stanford, CA.
Miller, M. and Baker, J. (2015). “Ground-motion intensity and damage map selection for proba-

26
 bilistic infrastructure network risk assessment using optimization.” Earthquake Engineering &
Structural Dynamics, 44(7), 1139–1156.
Miller-Hooks, E., Zhang, X., and Faturechi, R. (2012). “Measuring and maximizing resilience of
freight transportation networks.” Computers & Operations Research, 39(7), 1633–1643.
MR, H.-M. (2003). “Multi-hazard loss estimation methodology: Earthquake model.” Department
of Homeland Security, FEMA, Washington, DC, 235–260.
Nabian, M. A. and Meidani, H. (2018a). “Accelerating stochastic assessment of post-earthquake
transportation network connectivity via machine-learning-based surrogates.” Transportation Re-
search Board 97th Annual Meeting, Washington, DC.
Nabian, M. A. and Meidani, H. (2018b). “Deep learning for accelerated seismic reliability analysis
of transportation networks.” Computer-Aided Civil and Infrastructure Engineering, 33(6), 443–
458.
Olden, J. D., Joy, M. K., and Death, R. G. (2004). “An accurate comparison of methods for
quantifying variable importance in artificial neural networks using simulated data.” Ecological
modelling, 178(3-4), 389–397.
Özdamar, L. and Demir, O. (2012). “A hierarchical clustering and routing procedure for large scale
disaster relief logistics planning.” Transportation Research Part E: Logistics and Transportation
Review, 48(3), 591–602.
Ribeiro, M. T., Singh, S., and Guestrin, C. (2016). “"why should I trust you?": Explaining the
predictions of any classifier.” Proceedings of the 22nd ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016,
1135–1144.
Rokneddin, K., Ghosh, J., Dueñas-Osorio, L., and Padgett, J. E. (2013). “Bridge retrofit prioritisa-
tion for ageing transportation networks subject to seismic hazards.” Structure and Infrastructure
Engineering, 9(10), 1050–1066.
Tomar, A. and Burton, H. V. (2021). “Active learning method for risk assessment of distributed
infrastructure systems.” Computer-Aided Civil and Infrastructure Engineering, 36(4), 438–452.
Wald, D. J. and Allen, T. I. (2007). “Topographic slope as a proxy for seismic site conditions and
amplification.” Bulletin of the Seismological Society of America, 97(5), 1379–1395.
Werner, S. D., Taylor, C. E., Cho, S., Lavoie, J.-P., Huyck, C. K., Eitzel, C., Chung, H., and Eguchi,
R. T. (2006). “Redars 2 methodology and software for seismic risk analysis of highway systems.”
Report no.
Xie, Y., Ebad Sichani, M., Padgett, J. E., and DesRoches, R. (2020). “The promise of implementing
machine learning in earthquake engineering: A state-of-the-art review.” Earthquake Spectra,
36(4), 1769–1801.
Yu, H., Wu, Z., Wang, S., Wang, Y., and Ma, X. (2017). “Spatiotemporal recurrent convolutional
networks for traffic prediction in transportation networks.” Sensors, 17(7), 1501.

27