Machine Learning in Seismology

About Machine Learning

The application of machine learning in seismology is revolutionising our ability to analyse seismic signals. Machine learning algorithms have proven extremely effective in processing large amounts of seismological data. These algorithms can automatically analyse seismological data to differentiate seismic signals from background noise and detect the arrival times of different seismic phases, such as P-waves and S-waves, providing precise information about earthquake origins and magnitudes. These methods not only improve the speed and accuracy of earthquake detection but also minimise human bias in analysis. For sure, when there are only a few stations in the network, one person can easily analyse the data. However, when the network consists of several dozen stations, the traditional approach can be very time-consuming and must be performed by multiple people. Each analyst then introduces their own bias in signal detection, so the idea of having a computer perform this task with uniform uncertainty is very appealing. The result is a stronger and more efficient seismic monitoring system, crucial for early warning systems and hazard assessment. Using the example of a local earthquake recorded at the CRONOS (9H) and Du-Net (DN) network stations, we will explain some of the advantages and (current) disadvantages of using machine learning in seismology.

About Machine Learning Algorithms

In earthquake analysis using machine learning, several key factors are important for successful results. First and foremost, access to high-quality seismic data is crucial. The quantity and quality of data directly affect the performance of machine learning models. With our densified network, we can say that this condition is met. Additionally, choosing the best algorithms and, in particular, training data sets for these algorithms significantly impacts the model’s ability to accurately detect seismic events and phases. In this analysis, we will use the machine learning model EQTransformer (Mousavi et al., 2021), which is trained on the INSTANCE (Michelini et al., 2021) dataset. The INSTANCE dataset contains 1,159,249 seismograms from 54,008 local earthquakes (epicentral distance < 300 km) and 132,330 records of background noise recorded at 620 stations in Italy.

EQTransformer uses a so-called “Transformer” neural network, a type of artificial intelligence that is very good at recognizing patterns in data. Imagine looking at a complex series of waves on a screen – some of these signals are just background noise (like from traffic or wind), but others are actual seismic waves from earthquakes. EQTransformer is trained to differentiate these types of signals. Before it can do this, EQTransformer trains with many examples of seismic data (in this case, the INSTANCE dataset). Using this training dataset, which has been labelled and analysed manually, it learns what the characteristic recordings of earthquakes look like. Then, while analysing data, the algorithm looks for specific features that indicate an earthquake. These characteristics are based on the shapes, sizes, and times of waves in seismic data (Figure 1).

Figure 1. An example of how EQTransformer pays attention to different aspects of seismological data. The neural network, with its various layers, focuses on different parts of the earthquake record, thus defining what is an earthquake (a) and when the P phase (b) and S phase (c) begin. Figure adopted from Mousavi et al. (2020).

Example Analysis

As an example, we use a recent earthquake of local magnitude ML4.3 with an epicentre 10 km southwest of Metković at a depth of 4 km. Figure 2. shows the vertical component of ground motion velocity recorded at CRONOS and Du-Net network stations.

Figure 2. (left) Map of the temporary seismological networks CRONOS (blue triangles) and DuFault project (green triangles). The red star marks the epicentre of the earthquake. (right) Recording of the vertical component of ground motion velocity of the earthquake with an epicentre 10 km southwest of Metković that occurred at 17:01 on September 23, 2023, at the seismic stations of the CRONOS project (HRXX label) in Northern and Central Dalmatia and the DuFAULT project (Du-Net network, DFXX label) in the wider Dubrovnik area. Stations are arranged by epicentral distance.

To assess the success of earthquake phase detection, an analysis of this earthquake was done in the traditional way – by manually detecting the arrivals of P and S waves. Our experienced analyst performed this task, and her detections for 4 stations (HR10A, HR12A, DF02, and DF06) can be seen in the following figure (Figure 3).

Figure 3. Detection of P and S waves for 4 stations (HR10A, HR12A, DF06, and DF02) using the manual method. The arrival time of the P wave is marked in light blue, and the arrival time of the S wave is marked in orange.

Now that we have manually determined the arrival times of the P and S phases for all stations used in the analysis, we can run automatic detectors and see how they perform this task.

Figure 4. Comparison of the manual method and automatic earthquake detection using EQTransformer.

Figure 4. shows the automatic detection of earthquakes and seismic phases and compares it with the manual method for four selected seismic stations near the epicentre. The image for each station is divided into two parts. The upper image shows the seismogram with phase arrivals marked by the manual method (light blue and orange lines) and using EQTransformer (blue and red lines). The lower part of the image shows how much weight, i.e., probability, EQTransformer gives to its results. The blue line indicates what EQTransformer considers an earthquake, the green curve the time and probability of P-wave arrival, and the red curve the S-wave. Here we already see one advantage of this method, which is the quantified certainty of choosing the arrival time of each wave.

As mentioned in the introduction, in the case of a series of earthquakes recorded at dozens of stations, detecting earthquakes and different phases is a laborious and lengthy process, especially if the entire analysis is done by one analyst. Therefore, a whole team of people usually performs this task, a team with individuals of different work experiences. An experienced seismologist may analyse their data perfectly, while a less experienced one may do their part somewhat worse. In the final earthquake catalogue, all phase arrival time choices have equal weight, and the end user of these data does not know what uncertainties were involved in choosing these times. Machine learning methods allow us this quantification of certainty in phase selection.

The main question is whether machine learning methods can already replace manual methods. As promised in the introduction, here we come to the (current) disadvantages of these methods. Figure 5 shows the vertical component of station HR10A. The figure is an enlarged part of the seismogram around the arrival time of the P wave, and we see the difference between the selection by the manual method and the EQTransformer selection. Two problems are visible in this figure. The first, the obvious one, is the difference of 0.2 s between the two choices. This might not sound like a big difference, but considering that the P wave (in the upper 10 km of the Earth’s crust) spreads at a speed of about 5 km/s, a difference in the choice of time of 0.2 s leads to an error in the location of the earthquake of 1 km. The second problem is visible in Figure 5b, which is the polarity of the first shift (selected) of the P wave. The direction of the first shift of the P-wave is used to determine the focal mechanism which describes the earthquake source, i.e., shifts on the fault. This is one of our insights into the earthquake source itself, and it is extremely important to accurately determine the polarity of the first shift. In this example, we cannot be sure whether this first shift is upwards or downwards.

Figure 5. Zoomed in arrival of the P-wave on the vertical component of station HR10A.

The answer to whether machine learning methods can already confidently replace traditional methods is – not yet. However, if we continue to train our models with updated data and create local/regional datasets for training models, this answer could change in the near future.


Michelini, A., Cianetti, S., Gaviano, S., Giunchi, C., Jozinović, D., & Lauciani, V. (2021). INSTANCE – The Italian Seismic Dataset For Machine Learning. Istituto Nazionale di Geofisica e Vulcanologia (INGV).

Mousavi, S.M., Ellsworth, W.L., Zhu, W., Chuang, L.Y., Beroza, G.C., “Earthquake Transformer: An Attentive Deep-learning Model for Simultaneous Earthquake Detection and Phase Picking “. Nature Communications, (2020).

Prepared by Dr. Dinko Šindija