Developing predictive models of earthquake induced landslides using advanced neural networks

Background

Earthquake triggered landslides are a significant contributor to fatalities in large magnitude earthquakes. In addition to the immediate deaths and injuries caused by landslides the disruption to road infrastructure results in challenges in delivering aid and managing search and rescue operations. Generally, earthquakes bigger than M=5.0 can trigger landslides ranging from single block landslides (M=5.0+) to large rock and debris avalanches (M=6.5+) which can travel considerable distances. In spite of the hazard posed by this secondary earthquake effect, the controlling factors remain poorly understood, partly because of the lack of robust seismological data at the site of the landslide.

Traditionally, for the analysis of the stability of slopes subject to earthquake shaking, the ground motions have been characterised as a peak horizontal ground acceleration, a seismic coefficient (for limit-state, pseudo-static analyses) or an acceleration-time history (finite or discrete element methods). For sites that are distant from the earthquake source, such approaches yield reasonable results, however, there are considerable uncertainties in ground motion prediction when the distance between the seismic source and site of concern is relatively small (c. 10-20 km). This makes the reliability of ground motion predictions, and hence assessment of earthquake triggered landslide hazard, unreliable. Additionally, other factors such as topographic amplification, directionality effects and frequency content/wavelength of the waves inducing slope movement also contribute to problems of assessing stability of slopes.

General approach

The 2016 Central Italy seismic sequence, composed of three main earthquakes with epicentres in Amatrice (Mw=6.0), in Castelsantangelo sul Nera (Mw=5.9), and in Norcia (Mw=6.5) resulted in considerable landslide activity in this mountainous region. The landslide effects were extensively catalogued into the Italian Catalogue of Earthquake-Induced Ground Failures (CEDIT), and a good database of strong motion data were collected on both permanently installed instruments and temporary accelerometers. There was, however, poor correspondence between the location of the instruments and the ground failures,  resulting in a difficulty in parameterization of the seismic motion controls at the points where ground effects were triggered. However, the high quality of the data collected during this seismic sequence for both the earthquakes and the induced ground effects allows the opportunity to explore the factors controlling landslide type and occurrence during strong shaking.

In order to understand the controlling factors triggering landslides in the near-fault area sufficiently well to use such controls as predictive tools, we are going to reach the following objectives.

1) A selection of landslides (for example see the image on the left) spanning the range of ground failure types will be selected in CEDIT database and the a series of conceptual ground models will be developed for these sites. Appropriate mechanical properties will be assigned based on geological and geotechnical characterisation of the sites.

2) Finite difference modelling will be used to generate site specific shaking data for the selected ground failures. This approach will be validated against ground motion records recorded throughout the area held in databases at the National Institute of Geophysics and Volcanology and at the Italian Civil Protection databases. Ground motion parameters will include standard parameters such as ground acceleration but would also include wave frequency, duration of shaking, topographic response and directionality effects.

3) Geospatial databases will be developed for ground failure features, strong motion patterns and other landslide forcing parameters for the area affected by the earthquake sequence.

4) An artificial neural network such as Multilayer perceptron neural network or Deep learning approaches as a way of learning complex patterns based on the input layers of data. These non-linear models will allow the relationships between input data to be examined and the critical predictive parameters will be identified.

5) the predictors developed using the approaches developed in (4) will be tested against other events such as the Mw=6.3 L’Aquila Earthquake or the Mw=6.0 Umbria and Marche Earthquake.

The ultimate aim of understanding the critical parameters which controlled the initiation of landslides in the nearfield of seismic sources is to develop predictive tools for these areas and start the discovery of better tools for the evaluation of the hazards posed by earthquake triggered landslides more generally.

About the applicant

The successful candidate would be expected to spend at least one month of each year working with research partners at the Sapienza University of Rome and the University of Cassino and Southern Lazio. This may involve geological fieldwork and data analysis. Applicants will have the opportunity to attend subject specific elements of the MSc Engineering Geology such as rock engineering. Applicants will expect to demonstrate good levels of numeracy. Prior experience of numerical modelling or engineering geology is advantageous but not a prerequisite.

References:

https://doi.org/10.1007/s10346-019-01162-2

https://doi.org/10.1007/s10064-023-03163-x