Earthquake Magnitude Prediction in Chile Using

2022 IEEE 7th International Conference on Recent Advances and Innovations in Engineering (ICRAIE) | 978-1-6654-8910-2/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICRAIE56454.2022.10054323

Neural Network
Rohit K Bharadwaj Sumanta Pasari
Department of CSIS & Department of Mathematics Department of Mathematics
Birla Institute of Technology and Science Pilani Birla Institute of Technology and Science Pilani
Pilani, India Pilani, India
[email protected] [email protected]

Sonu Devi
Department of Mathematics
Birla Institute of Technology and Science Pilani
Pilani, India
[email protected]

Abstract—In this study, we implement an earthquake magni- Various earthquake precursors have been used in the past
tude prediction model using a neural network for a test region in for earthquake prediction. Few of them include unusual animal
Chile. For this, the epicenter of earthquake is located on a mesh behaviour [5], electromagnetic precursors [6] and changes in
with dimensions of 1°×1°. We adopt a zonation scheme originally
proposed by Reyes and Cardenas [1]. The scheme uses increments groundwater levels [7]. Previous studies that have focused
in b−value and other input parameters to incorporate G-R linear on magnitude prediction are commonly based on past earth-
relation and Bath’s law. The model enables the prediction of the quake magnitude data in an area of interest. The magnitude-
maximum magnitude for a given cell within the next five days. frequency Gutenberg-Richter (G-R) inverse power law [8]
Common seismological parameters are used for the performance has been extensively used. This law states that earthquake
evaluation of the model. Results show satisfactory performance
of the proposed model in comparison to other existing models. magnitudes follow an exponential model. The total event count
Index Terms—earthquake prediction, neural network, time N having magnitude M or higher is related to the magnitude
series M as follows:
log10 N = a − bM (1)
I. I NTRODUCTION In Equation 1, the constant parameter b is often called as
Reliable prediction of earthquake occurrence in terms of the b−value, which is approximately 1 for seismogenic areas.
the epicentral location, time, and magnitude is one of the The other parameter a represents the level of background
potential research areas in Seismology. The topic of earthquake seismicity. Out of these two parameters, the b−value is more
prediction has been quite controversial, as some notable sci- important and it has been used extensively to capture patterns
entists believe that the exact size of a future event, occurrence in earthquake studies. The b−value analysis helps model
time, or the exact location of a single impending earthquake the seismicity and physical characteristics of the area under
can not be predicted [2]. In contrast, many researchers have analysis [9]. Recently, b−values were also considered by
suggested various physical and empirical methods for earth- Gulia and Weimer [10] to discriminate between earthquake
quake prediction in their respective study regions [3], [4]. foreshocks and aftershocks. Here we have utilized various
Moreover, for earthquake-resistant engineering design, disaster parameters based on b− value and the G-R law.
management, catastrophic insurance, and seismic awareness After a general introduction in this section, Section II
programs, earthquake prediction or its associated concept, presents a few important works on earthquake prediction
earthquake forecasting, has numerous applications. using machine learning algorithms. Section III describes the
As on today, there is no successful model that accurately methodology of the study, whereas Section IV explains and
predicts earthquakes in a given region of interest. However, interprets the results obtained through the study.
the last few decades have observed significant developments II. R ELATED W ORKS
in earthquake prediction tools and methods in terms of highly The problem of earthquake prediction is challenging. De-
precise instrumentation, accurate data acquisition techniques, spite the great efforts made by scientists all over the world, no
space-based geodesy, and numerical simulations. In addition, method has been found yet to be successful in tackling this
the efficient use of several data-driven methods, machine learn- issue reliably [11]. The randomness involved in earthquake
ing, and artificial intelligence techniques have demonstrated generation or the highly complex nonlinear process probably
their performance in seismological applications. makes it difficult to predict earthquake magnitude, location,
978-1-6654-8910-2/22/$31.00 © 2022 IEEE and the exact time of the future event(s).

294
Authorized licensed use limited to: BIRLA INSTITUTE OF TECHNOLOGY AND SCIENCE. Downloaded on November 30,2024 at 12:35:41 UTC from IEEE Xplore. Restrictions apply.
 Many machine learning models of earthquake prediction 1) First five inputs for the neural network: The increments
have been proposed in the past. For instance, Panakkat and of b−value are calculated from the following expressions.
Adeli [12] used three models of the neural network, namely the
artificial neural network (ANN), recurrent neural network, and
radial basis function neural network to carry out earthquake ∆b1i = bi − bi−4 ≡ x1i (3)
prediction. They used various seismicity indicators for this ∆b2i = bi−4 − bi−8 ≡ x2i (4)
purpose. They have also considered some parameters based ∆b3i = bi−8 − bi−12 ≡ x3i (5)
on characteristic earthquake magnitude distribution along with
G-R law. Similarly, Reyes et al. [13] have considered ANN- ∆b4i = bi−12 − bi−16 ≡ x4i (6)
based models for earthquake prediction in Chile. Using Bath’s ∆b5i = bi−16 − bi−20 ≡ x5i (7)
Law, Omori-Utsu’s law, and b−value based parameters, their
model predicts the chance of an event larger than a threshold From the above equations, it is clear that we require at least
magnitude within a five-day duration in a given area. Although 70 events to calculate these five inputs for the neural network.
the methodology is impressive, the authors have not used any 2) Sixth Input: The maximum recorded magnitude during
state-of-the-art neural network model in their approach. the previous week is the sixth input variable, x6i . With this
In light of the above discussion, the present work proposes input, we are essentially providing the neural network the
a neural network architecture with seven input parameters required information to use Bath’s law [18].
for a reliable earthquake magnitude prediction. Particularly,
we consider the increments in b−value as one of the inputs, x6i = max Ms (8)
t∈[−7,0)
to assume that the model would incorporate G-R law while
training. We formulate a classification problem through a Here, t denotes the time in days.
supervised machine learning technique. The functionality of 3) Seventh Input: For the final input, x7i , we consider the
the model will be evaluated based on precision and accuracy the occurrence probability of an event larger than a reference
measures, as explained in a later section. threshold, say Mw = 6, to capture dynamic G-R relationship.
The associated expression is as follows:
III. M ETHODOLOGY
3bi
A. Data Acquisition x7i = P (Mw ≥ 6.0) = e− log e = 10−3bi (9)
The data for the present study is downloaded from the
public IRIS website (http://ds.iris.edu/ieb/) that uses the global C. Output Variable
instrumental earthquake catalog provided by the International The output variable, yi , is the predicted maximum mag-
Seismological Centre (ISC) [14]. We use the minimum magni- nitude Mw in the next five days. We set yi to zero for
tude as 3.0, dates from 2002-03-01 to 2015-08-31, and a region cases where the predicted magnitude is below the threshold
bounded by 710 W − 720 W and 350 S − 360 S to download the magnitude M0 = 3. Formally,
earthquake data in an XML format. The study site corresponds
to the Talca city of Chile. The location was chosen following yi = max Mw , (10)
the work by [13], so that the results can be compared. t∈(0,5]

B. Inputs for Neural Network where the time t is measured in days.

The proposed neural network requires seven input param-
eters to predict a scalar value. Before we compute these D. Neural Network Architecture
parameters, the variable b−value is calculated through the We consider seven input neurons in the architecture. Fol-
below equation [15]. lowing [19], we consider one hidden layer and we apply Kol-
mogorov’s theorem [20] to decide the hidden layer’s neuron
log(e)
bi = P49 (2) number. Based on this theorem, we require (2n + 1) neurons.
1
50 j=0 Mi−j − M0 With n = 7, the hidden layer has 15 neurons. The number of
In the above equation, the magnitude for the i-th earthquake is synaptic connections of neural network equals to
denoted by Mi , whereas M0 = 3 is the threshold magnitude. [input neurons][hidden layer neurons] + [hidden layer neu-
We have used the magnitude information of the last 50 rons][no of hidden layers] + 2, which is 122.
earthquakes to calculate the b−value. Therefore, we require at least 122 linearly independent
Many studies have found that there is an increase in training data points to train the proposed neural network
b−value following large events and b−value reduces before the model.
aftershock [16], [17]. Therefore, variation in b−values plays The output contains only one neuron that uses the sigmoid
a crucial role in determining the occurrence of an earthquake activation function. The neural network is triggered when the
in a spatial region. Here, the first five inputs (x1i to x5i ) of output is greater than a given threshold. This threshold is
the proposed network comprise the variation of b−values, as obtained by adding the mean magnitude of 122 training data
highlighted below. points to 0.6 times the standard deviation.

295
Authorized licensed use limited to: BIRLA INSTITUTE OF TECHNOLOGY AND SCIENCE. Downloaded on November 30,2024 at 12:35:41 UTC from IEEE Xplore. Restrictions apply.
 E. Network Training TABLE I
P ERFORMANCE M ETRICS FOR THE TRAIN SET.
To train the ANN, we make use of Keras library [21], and
train the ANN for 3500 epochs using standard backpropaga- Parameters Value
tion algorithm [22]. We have used Adam optimizer [23] with TP 11
TN 92
learning rate, lr = 1 × 10−2 , and the default values for the FP 6
other parameters. FN 13
P0 87.6%
IV. R ESULTS P1 64.7%
Sp 45.8%
A. Performance Evaluation Sn 93.8%
Average 72.9%
To evaluate the functionality of the proposed neural net-
work, the following measures have been used.
1) True positives (TP): It is the cumulative count when the
predicted earthquake size crosses the threshold during
the next 5 days, given that the actual earthquake size
also crosses the threshold value.
2) True negatives (TN): It is the cumulative count when
the predicted earthquake size is lesser than the threshold
during the next 5 days, given that the actual earthquake
size is also lesser than the threshold value.
3) False positives (FP): It is the cumulative count when the
predicted earthquake size crosses the threshold during
the next 5 days, given that the actual earthquake size is
lesser than the threshold value.
4) False negatives (FN): It is the cumulative count when Fig. 1. Validation loss is shown in red, whereas the training loss is shown in
the predicted earthquake size is lesser than the threshold blue.
during the next 5 days, given that the actual event size
is larger than the threshold value.
C. Test Set
Along with the above metrics, we have used the following
two precision measures, P0 and P1 , to understand how well The test set comprises 481 earthquake data spanning from
the model performs when the prediction is negative or positive. April 2004 to August 2015. The error measures for the test
The P0 and P1 are calculated as follows [13]: data are summarized in Table II.
Based on the results provided in Table I and Table II, we
TN observe that the proposed model has better performance than
P0 = (11)
TN + FN the model by Reyes et al. [13]. The current model has obtained
TP 72.9% overall accuracy on the training set, whereas the model
P1 = (12)
TP + FP by Reyes et al. [13] obtained 70.3% accuracy on the train set.
Additionally, we have two other measures to judge the net- In addition, the present model yields 61.0% overall accuracy
work’s performance. These are commonly known as sensitivity on the test dataset which is significantly higher than the 49.9%
(true positive rate, recall or hit rate; Sp ) and specificity (true accuracy obtained in [13].
negative rate, Sn ). It may be noted that the test data set used by Reyes et al.
[13] comprises only 45 events, whereas the present model had
TP 481 events in the test data. Thus, it appears that the current
Sp = (13) model has higher generalizing capability along with better
TP + FN
TN performance. For the training set, the current model has only
Sn = (14) 13 false negatives, whereas the model in [13] obtained 18 false
TN + FP
negatives. Low values of false negatives are always preferred
B. Train Set in earthquake prediction because of the high risk associated
In the training set, 122 earthquake data spanning from with earthquakes and their impact on human lives and society.
March 2002 to April 2004 is used. The data was normalized
before passing it to the neural network to improve the training V. S UMMARY
accuracy. The neural network has been trained for 3500 epochs The present work implements an earthquake magnitude
with the mean squared error (MSE) loss function. Table I prediction model using a neural network. The initial zonation
summarizes the values of different error measures. An R2 scheme and the initial set of parameters are adopted from
value of 0.711 is obtained on the train set. The validation Reyes and Cardenas [1] and Reyes et al. [13]. For the dataset
loss and the training loss are collectively plotted in Fig 1. in a test region in Chile, we find that the proposed scheme

296
Authorized licensed use limited to: BIRLA INSTITUTE OF TECHNOLOGY AND SCIENCE. Downloaded on November 30,2024 at 12:35:41 UTC from IEEE Xplore. Restrictions apply.
 TABLE II [15] P. Nuannin, “The potential of b-value variations as earthquake precur-
P ERFORMANCE M ETRICS FOR THE TEST SET. sors for small and large events,” Ph.D. dissertation, Acta Universitatis
Upsaliensis, 2006.
Parameters Value [16] S. J. Gibowicz, “Frequency-magnitude, depth, and time relations for
TP 62 earthquakes in an island arc: North island, New Zealand,” Tectono-
TN 253 physics, vol. 23, no. 3, pp. 283–297, 1974.
FP 121 [17] S. Wiemer, M. Gerstenberger, and E. Hauksson, “Properties of the
FN 45 aftershock sequence of the 1999 m w 7.1 hector mine earthquake:
P0 84.9% implications for aftershock hazard,” Bulletin of the Seismological Society
P1 33.8% of America, vol. 92, no. 4, pp. 1227–1240, 2002.
Sp 57.9% [18] M. Båth, “Lateral inhomogeneities of the upper mantle,” Tectonophysics,
Sn 67.6% vol. 2, no. 6, pp. 483–514, 1965.
Average 61.0% [19] P. Perez and J. Reyes, “An integrated neural network model for pm10
forecasting,” Atmospheric Environment, vol. 40, no. 16, pp. 2845–2851,
2006.
[20] R. Hecht-Nielsen, “Kolmogorov’s mapping neural network existence
has better performance than Reyes et al. [13]. Nevertheless, theorem,” in Proceedings of the International Conference on Neural
Networks, vol. 3. IEEE Press New York, 1987, pp. 11–14.
there remains a lot of scope in the future to expand the current [21] F. Chollet et al. (2015) Keras. [Online]. Available:
work. For example, one may use deep learning architecture to https://github.com/fchollet/keras
improve the model accuracy. [22] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning internal
representations by error propagation,” California Univ San Diego La
Jolla Inst for Cognitive Science, Tech. Rep., 1985.
ACKNOWLEDGMENTS [23] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
2014. [Online]. Available: https://arxiv.org/abs/1412.6980
Partial financial support was provided by IRDR-ICoE Tai-
wan and DST-SERB (MATRICS scheme). The third author
acknowledges financial support from BITS Pilani in terms of
PhD fellowship.

R EFERENCES
[1] J. Reyes and V. H. Cárdenas, “A Chilean seismic regionalization through
a kohonen neural network,” Neural Computing and Applications, vol. 19,
no. 7, pp. 1081–1087, 2010.
[2] R. J. Geller, D. D. Jackson, Y. Y. Kagan, and F. Mulargia, “Earthquakes
cannot be predicted,” Science, vol. 275, no. 5306, pp. 1616–1616, 1997.
[3] D. J. Brehm and L. W. Braile, “Intermediate-term earthquake prediction
using the modified time-to-failure method in Southern California,”
Bulletin of the Seismological Society of America, vol. 89, no. 1, pp.
275–293, 1999.
[4] M. V. Matthews, W. L. Ellsworth, and P. A. Reasenberg, “A Brownian
model for recurrent earthquakes,” Bulletin of the Seismological Society
of America, vol. 92, no. 6, pp. 2233–2250, 2002.
[5] J. L. Kirschvink, “Earthquake prediction by animals: Evolution and
sensory perception,” Bulletin of the Seismological Society of America,
vol. 90, no. 2, p. 312–323, Jan 2000.
[6] P. Varotsos, “Seismic electric currents,” Prakt. Akad. Athenon, vol. 56,
pp. 277–286, 1981.
[7] I. Kissin and A. Grinevsky, “Main features of hydrogeodynamic earth-
quake precursors,” Tectonophysics, vol. 178, no. 2-4, pp. 277–286, 1990.
[8] B. Gutenberg and C. F. Richter, “Earthquake magnitude, intensity, en-
ergy, and acceleration,” Bulletin of the Seismological Society of America,
vol. 32, no. 3, pp. 163–191, 1942.
[9] K. Lee and W.-S. Yang, “Historical seismicity of Korea,” Bulletin of the
Seismological Society of America, vol. 96, no. 3, pp. 846–855, 2006.
[10] L. Gulia and S. Wiemer, “Real-time discrimination of earthquake
foreshocks and aftershocks,” Nature, vol. 574, no. 7777, pp. 193–199,
2019.
[11] K. F. Tiampo and R. Shcherbakov, “Seismicity-based earthquake fore-
casting techniques: Ten years of progress,” Tectonophysics, vol. 522, pp.
89–121, 2012.
[12] A. Panakkat and H. Adeli, “Neural network models for earthquake
magnitude prediction using multiple seismicity indicators,” International
Journal of Neural Systems, vol. 17, no. 01, pp. 13–33, 2007.
[13] J. Reyes, A. Morales-Esteban, and F. Martı́nez-Álvarez, “Neural net-
works to predict earthquakes in Chile,” Applied Soft Computing, vol. 13,
no. 2, pp. 1314–1328, 2013.
[14] D. A. Storchak, D. Di Giacomo, I. Bondár, E. R. Engdahl, J. Harris,
W. H. K. Lee, A. Villaseñor, and P. Bormann, “Public Release of the
ISC–GEM Global Instrumental Earthquake Catalogue (1900–2009),”
Seismological Research Letters, vol. 84, no. 5, pp. 810–815, 09 2013.
[Online]. Available: https://doi.org/10.1785/0220130034

297
Authorized licensed use limited to: BIRLA INSTITUTE OF TECHNOLOGY AND SCIENCE. Downloaded on November 30,2024 at 12:35:41 UTC from IEEE Xplore. Restrictions apply.