Earthquake Probabilistic Forecasting and Detection using Neural Networks


Impending rupture in man-made structures, natural landslides or earthquake faults is sometimes preceded by slow preparatory strain and increased acoustic emissions. Modern broad-band seismic data records both the very short-lived (seconds to minutes), dynamic rupture instability events and the long-lived (hours), slow deformation instability events during the preparatory or nucleation phase. These processes emit low amplitude signals which can be hidden in plain sight within noisy time series, and difficult to detect with traditional methods. In addition, the precise nature of these emissions and their generating process remain unclear.
Recently, machine learning (ML) and specifically convolutional neural networks (CNNs), have shown great potential in revealing patterns hidden in the noise. A few examples of successful seismological applications are opening interesting research avenue. In addition to seismological data from natural events, laboratory experiments performed on rocks provide an opportunity to deepen our exploration of the dynamics of rupture onset, under a controlled and well-monitored environment. A CNN prototype has been tested at Durham University and has shown promising results in detection and forecasting of both natural earthquakes (a few hours in advance of magnitude 6 events off the coast of Japan, Fig 1-2) and ruptures in laboratory experimental rock samples.
However, the accuracy of the CNN method in locating and in estimating the magnitude of the future event is yet unknown. In addition, it was not tested in many different tectonic environments.
The aim of this PhD is to develop and generalise this CNN prototype for the analysis of seismic signals, to test its generality, and explore the geophysical origin of the detected signal in the framework of the seismic cycle. In addition, the student will investigate the potential integration of the ML in (1) seismic Early Warning protocols, (2) real-time scenarios of probabilistic forecasting and risk mitigation, (3) automation and improvement in the detection, location and classification of seismic events in regional seismic catalogues (it has been shown that up to 60% more events can be detected by using CNN methods).

Click on an image to expand

Image Captions

Figure 1: Mw ≥ 6 earthquakes investigated in the Japan region (time span 2012-2020) with the CNN method. Orange circles indicate the location of the earthquake epicentres. The yellow and green triangles indicates the location of the seismic stations used in the forecasting.

Figure 2: Changes in the average test accuracy over different time intervals labelled as precursors by the CNN method. The CNN was trained and tested using the set of Japanese earthquakes and the seismic broadband station shown in Fig. 1. The red, dashed, vertical line indicates the start of a significant increase (above the standard deviation) in the test accuracy.


We propose to expand, generalise and systematically test our current prototype of Convolutional Neural Network with time sequences of natural events in the vicinity of earthquake faults (borehole strain-meters, broad-band seismic stations, GPS data) and on laboratory experimental data of rock sample rupture. Most promising result in classifying pre-earthquake time windows have been obtained by integrating in Long-Short-Term-Memory in our prototype of Fully Convolutional Networks (LSTM-FCN, Karim et al., 2017) for the analysis of time series. A wider variety of algorithms can be tested to investigate which methods are best suitable for each specific problem (e.g, forecasting, classification, detection).
At the moment the prototype is developed in the Python programming environment. One advantage of Python is the great number of ready made, open-source and free modules for artificial intelligence that reduce enormously the development time. In particular, our prototype is based on the Keras module libraries.

Project Timeline

Year 1

The initial months of the projects will be mainly devoted to an in-depth study of earthquake seismology, fault mechanics, and state-of-the-art machine learning techniques. The student will conduct a Bibliographic research and receive training in programming tools (python, machine learning, data analysis). The student will also become acquainted with protocols and techniques of rock mechanics laboratory experiments.
My the mid- to end-of-year, a first series of broad-band seismic network data from public databases (IRIS, EPOS) will be sourced, focussing on earthquake sequences on major intra-plate and inter-plate faults. Short duration (minutes to hours) laboratory experiments will be conducted on experimental earthquake faults. The existing CNN will be trained and tested on these data, and its efficiency will be compared with the results obtained on the Japanese earthquake series, to evaluate the portabilty and generality of the method to different tectonic environments.

Year 2

The second year will be devoted to the design and systematic testing of alternative version of the CNN algorithm, and to broaden the database on which the algorithms are being tested. The most accurate and efficient type of algorithms will be selected, and expanded into a full analytic machine-learning suite. The accuracy will be cast in terms of probability of forecasting successfully within a given time, space and magnitude window, versus over-prediction (cry wolf effect). The data will be analysed to determine what type of signal or combination of signals triggers the detection of a precursor. This will be interpreted in terms of a changes in the slip dynamics in the fault system in the preparatory phase of an earthquake, and integrated in a model of the earthquake cycle. Time should be allowed to finalise the publication and dissemination of main results, and to the write-up of the PhD manuscript.

Year 3

The work started in the second year will be pursued, with the additional testing of the developed analytical tools. The practical implementation of the algorithm as a possible probabilistic forecasting tool for earthquake rupture will be investigated. Automatic download of waveform data and meta data from a regional network in a target active seismic area will be performed using query software (e.g. ObsPyLoad) and fed into a real-time, continuously running 24h algorithm, with an aim at the possible integration in Earthquake Early Warning protocols. The forecasting accuracy algorithm will also be tested on data from in induced-seismicity sites (e.g., borehole injection operations for geothermal energy), with possible implication for the current “traffic-light” safety systems. Processing of regional seismic data will also be implemented with the aim to augment the number of detected events and provide an estimate of magnitude and location. magnitude. The student and the supervisory team will start to organise the results obtained into publishable form, with the aim to disseminate them through articles in scientific journals, conferences, seminars, web pages.

Year 3.5

Time should be allowed to finalise and confirm the tests, to the publication and dissemination of main results, and to the write-up of the PhD manuscript.

& Skills

– Programming in Python
– Machine learning applications
– Data analysis and signal processing
– Hazard, statistics and probabilistic forecasting
– Earthquake fault mechanics

References & further reading

– Galea, A. and Capelo, L. Applied Deep Learning with Python: Use scikit-learn, TensorFlow, and Keras to create intelligent systems and machine learning solutions. Packt Publishing, 2018.
– Goyal, P. and Pandey, S. and Jain, K. Deep Learning for Natural Language Processing: Creating Neural Networks with Python. Apress 2018.
– Guerin-Marthe, S., Nielsen, S., Bird, R., Giani, S. & Di Toro, G. Earthquake Nucleation Size: Evidence of Loading Rate Dependence in Laboratory Faults. Journal of Geophysical Research: Solid Earth 124(1): 698-708. 2019
– Hulbert, C. et al. Similarity of fast and slow earthquakes illuminated by machine learning. Nature Geoscience 12: 69-74. 2019
– Karim, F. et al. LSTM fully convolutional networks for time series classification, IEEE access 6, 1662-1669. 2017.
– Pattanayak, S. Pro Deep Learning with TensorFlow: A Mathematical Approach to Advanced Artificial Intelligence in Python. Apress, 2017.
– Rouet-Leduc, B., Hulbert, C. & Johnson, P. A. (2019), “Continuous chatter of the cascadia subduction zone revealed by machine learning”, Nature Geoscience 12(1), 75-79.

Further Information

Apply Now