[S-SS04] New trends in data acquisition, analysis and interpretation of seismicity

[SSS04-09] A Deep Learning Paradigm for Earthquake Hazard Assessment

*Haritha Chandriyan¹, Debjeet Mondal², Paresh Nath Singha Roy¹ (1.Department of Geology and Geophysics, Indian Institute of Technology Kharagpur, India., 2.CMPDI, Coal India Ltd, India)

Keywords:Correlation fractal dimension, Deep Learning, Multilayer Perceptron, Earthquake hazard

Fractals are intricate, self-replicating geometric patterns that exhibit similar structures at various levels of magnification. The fractal nature of earthquakes allows the determination of the Correlation Fractal Dimension (Dc), a metric for spatial clustering. In a prior study, numerical precursors were identified before significant events (Mw > 7), with values falling below 0.2 in Southern and Baja California (Chandriyan et al., 2022). Furthermore, the 2023 Turkey-Syria earthquake doublet (Mw 7.7 and 7.6) was found to be preceded by a Dc value of 0.16 (Chandriyan et al., 2024). This illustrates the potential for implementing hazard monitoring through Dc fluctuations. Consequently, we aim to develop an earthquake hazard forecasting method by incorporating deep learning (DL) techniques. Because of its ability to discern complex patterns and relationships in extensive datasets, DL enables researchers to uncover nuanced insights that may be challenging for traditional methods. In this context, Dc calculations were performed based on relative clustering, unlike computations considering a fixed point for the correlation integral estimation. Nearly 30 years of seismicity data have been assessed to study the variation of Dc in the California region. The entire earthquake dataset was segmented into multiwindows of 50 earthquakes in each set, and inter-earthquake distances for each of these sets were calculated to identify the trend in earthquake cluster distribution. Further classification of events was based on the frequency of their occurrence (ef) under a specified distance. This ef will be trained against the hazard status, which is classified into different levels based on Dc values, regarding the chance of the occurrence of M > 7 earthquakes in the study region. We will employ the multilayer perceptron (MLP) network for training and prediction purposes. It is categorized as a feedforward artificial neural network, characterized by its structure comprising multiple layers of nodes (neurons). MLPs can model complex non-linear relationships in data, making them suitable for tasks where simple linear models may not capture the underlying patterns effectively. Initial data collection, preliminary analyses, and classifications have been carried out. As the study is ongoing, results will be presented at the time of the conference presentation.