Seismological Oriented Machine lEarning (SOME) project

Abstract

Recent developments of machine learning (ML) algorithms and software-platforms (e.g. Keras, TensorFlow, PyTorch) have opened new frontiers for Earth sciences. In seismology, these advances have affected different aspects of the earthquake physics studies, such as ground motion prediction, seismic phase detection and identification, and seismic big-data analysis.Within the project Pianeta Dinamico (Working Earth) of Istituto Nazionale di Geofisica e Vulcanologia, funded by the Italian Ministry of University and Research, in 2021 we applied to an internal call with the aim of developing and using existing state-of-the-art machine learning techniques, and delivering useful benchmarking dataset for earthquake analysis. The project is named SOME (Seismological Oriented Machine lEarning).This multidisciplinary project tackles different tasks that highlight the potential and possible pitfalls of ML applications:Earthquake monitoring: testing and applying Convolutional Neural Network (CNN) and Graph Neural Network (GNN) architectures to predict the intensity measurements (IM) of medium-size seismic events (2.9 < M ≤ 5.1) recorded from a regional network. Seismic waveforms characterization: development of an unsupervised framework for hierarchical clustering of continuous data based on a deep scattering network (scatseisnet). A first application is aimed to detect and classify seismic data from a mainly aseismic region in NE Sardinia (Sos Enattos mining site) to assess the anthropogenic and natural noise levels. Development of a new picking algorithm: implementation of U-NET model architecture of PhaseNet algorithm by using characteristic functions derived from FilterPicker software. This newly developed software is called Domain Knowledge PhaseNet (DKPN). Creation of 2 ML dataset for earthquake studies: 1) INSTANCE dataset, containing the seismicity recorded between January 2005 and January 2020 by the national seismic network of INGV (~1.2 million three-component waveform traces), 2) AQUILA-2009 dataset containing the aftershock sequence of the 2009 Mw6.1  L’Aquila earthquake collected by a dense array of the permanent and temporary network deployed after the mainshock (>63,704 events, nearly >1.2 million 3C three-component traces). The INSTANCE and AQUILA-2009 dataset are already used as training sets for new picking algorithms, and will be employed for additional statistical analysis in the near future (e.g. hazard assessment, shakemaps) and transfer-learning approaches. The GNN for IM shows promising results for future developments for ground-shaking forecasting applications. The unsupervised learning clusterization algorithm clearly detects signals that differ from purely seismic ones, proving to be a great tool for seeking new patterns and features in time-series records. The DKPN algorithm achieves better results compared  to the original PhaseNet architecture, even if trained with a small dataset (<15.000 3C traces), and shows improved performance for cross-domain application.Overall, the SOME project has produced many deliverables, some of which have already been released. We also aimed to provide reproducibility of ML experiments, creating Docker applications suitable for ML-picking algorithms (e.g. EQ-Transformer, PhaseNet, GPD) and contributing to the improvement of existing libraries, like SeisBench, for benchmarking purposes. Indeed, reproducibility is an additional yet paramount issue that must be addressed by the seismological community when dealing with ML applications.