Manganiello, Ester (2023) Enhancing earthquake forecasting through innovative stochastic modelling. [Tesi di dottorato]

Anteprima

Testo manganiello_ester_35.pdf Download (7MB) | Anteprima

Abstract

The possibility to predict earthquakes has always been a hotly debated topic in the scientific community due to the destructive power of these events. Although we are not (yet?) able to predict deterministically with high precision where, when, and how big the next earthquake will struck, currently, several models are able to state, in probabilistic terms, that a seismic event can occur in a certain place, time interval and with a certain magnitude. These models have a big limitation, which consists of the inability to forecast large earthquakes with high probabilities. In fact, these models, such as ETAS model (Epidemic Type Aftershock Sequence), assume the independence of the magnitude: the magnitude of the next earthquake is randomly sampled from an exponential-like distribution, bounding the probability of large earthquake to the sampling of the right tail of such distribution. In this thesis, we aim to provide new insights on this problem that can be useful to improve the probabilistic forecasting models. Specifically, we analyze foreshocks, i.e., events that precede an event of higher magnitude: the goal is to understand if they have characteristics such as to be recognized a-priori (i.e., before the occurrence of a large seismic event), thus distinguishing them from the rest of seismicity. If so, their predictive power could be used to improve the probability of larger events. In this thesis, we use ETAS model as null hypothesis, i.e., able to represent all seismicity, and we compare the foreshock sequences of synthetic and real catalogues. The results highlight discrepancies between what the model is able to predict and what we observe in reality. However, these discrepancies do not concern all foreshock sequences, but only seismic sequences with peculiar characteristics: they are located mainly in areas of high heat flow, with a small/medium mainshock magnitude, and characterized by a high number of foreshocks. These sequences are commonly referred in the literature as swarms. Foreshock sequences occurring in low heat flux zone conform well with what expected by the ETAS model. To further investigate on the difference between earthquake sequences, we use a scalar parameter called ALD (Average Leaf Depth). This parameter is able to quantitatively represent the topological structure of the sequences, i.e., swarm type (with high ALD values) or burst type (the earthquake sequence explained by the ETAS model, with low ALD values). We test the null hypothesis that ETAS represents well the reality and therefore, that it is also able to reproduce seismic sequences with high ALD values (i.e., swarm). Comparing the ALD values of real sequences and synthetic sequences we obtained highly variable results depending on the specific ETAS model used and the type of source. This makes ALD a parameter to be used with extreme care and its variability makes it not so useful for the discretization of seismic sequences (i.e., between swarm and aftershock-type sequences). As in the literature there is no criterion to objectively differentiate the types of seismicity, we present a new methodology based on machine learning. First of all, we collect several information that characterize the seismic sequences in terms of space, time, and magnitude. We apply a Cluster Analysis to the collected dataset and identify the optimal number of clusters into which to divide the sequences. There are two predominant clusters that differ from a physical and statistical point of view: one group is mainly located in areas of high heat flow, is characterized by a high productivity and rate of foreshocks, therefore comparable to swarm-type sequences. Conversely, the other group has characteristics attributable to aftershock-type sequences. In conclusion, the above-mentioned method could be useful for recognizing the type of a seismic sequence (i.e., swarm or aftershock type), collecting all the information available on it (ALD, heat flow value corresponding to its location, productivity, etc.). If we were able to apply these methodologies in the short term and identify swarm-type sequences, we could reduce the probability that a large seismic event could occur imminently, thus improving the probability estimates of current models.

Downloads

Actions (login required)

Modifica documento