Deep Learning and Parallel Processing Spatio-Temporal Clustering Unveil New Ionian Distinct Seismic Zone

Antonios Konstantaras

doi:10.3390/informatics7040039

Abstract

This research work employs theoretical and empirical expert knowledge in constructing an agglomerative parallel processing algorithm that performs spatio-temporal clustering upon seismic data. This is made possible by exploiting the spatial and temporal sphere of influence of the main earthquakes solely, clustering seismic events into a number of fuzzy bordered, interactive and yet potentially distinct seismic zones. To evaluate whether the unveiled clusters indeed depict a distinct seismic zone, deep learning neural networks are deployed to map seismic energy release rates with time intervals between consecutive large earthquakes. Such a correlation fails should there be influence by neighboring seismic areas, hence casting the seismic region as non-distinct, or if the extent of the seismic zone has not been captured fully. For the deep learning neural network to depict such a correlation requires a steady seismic energy input flow. To address that the western area of the Hellenic seismic arc has been selected as a test case due to the nearly constant motion of the African plate that sinks beneath the Eurasian plate at a steady yearly rate. This causes a steady flow of strain energy stored in tectonic underground faults, i.e., the seismic energy storage elements; a partial release of which, when propagated all the way to the surface, casts as an earthquake. The results are complementary two-fold with the correlation between the energy release rates and the time interval amongst large earthquakes supporting the presence of a potential distinct seismic zone in the Ionian Sea and vice versa.

Highlights

Neural networks have long been established as effective global approximators [1,2]
The general result appears to be that the deep learning neural network managed to estimate time intervals between consecutive large earthquakes with variations in the range of a few months
Despite the presence of two unaccounted phenomena, i.e., the seismic clustering effect in 2008 and the appearance of no interim medium-large earthquake in between 2015 and 2018, the deep learning neural network was able to learn and to some extent account for the effect of the aforementioned information, an attribute due to deep learning in comparison with classical artificial neural networks that would have required the intervention of domain experts

Summary

Introduction

Neural networks have long been established as effective global approximators [1,2]. The main limitations observed were mainly due to their high dimensionality that required significant processing power and fast memory access times. Heterogeneous parallel processing [3] encompasses both short latency CPU cores for serial processes and large throughput GPU compute units for parallel processes, respectively It effectively resolves processing time and memory management issues as artificial neural networks are mostly organized in several layers comprised of multiple parallel neurons. This brought rise to deep learning [4] that invokes neural networks of extensive architecture that are capable of learning from the data thanks to their features’ extraction and classification capabilities. There is no possible way to measure the accumulated strain energy stored in tectonic underground faults [5]

Methods

Results

Discussion

Conclusion