Exploring the Potential of Deep Learning models in Focal Mechanism Computation: A Case Study from Switzerland

Abstract

The Swiss Seismological Service (SED) manages a dense seismic network in Switzerland that allows the detection and location of small magnitude earthquakes recorded on several stations. The current manual workflow involves picking of phase arrival times and P-wave first-motion polarities, as well as phase-type identification leading to the creation of a high-quality earthquake catalog and phase-pick dataset. Periodically, a dedicated seismologist conducts a thorough analysis of more significant earthquakes (usually ~ML>2.5) and computes focal mechanisms using P-phase polarities in combination with the HASH software. While this meticulous process guarantees the computation of reliable and high-quality focal mechanisms, it is time-consuming, since it requires a manual review for each earthquake.   In recent years, advancements in computational capabilities have empowered us to seamlessly integrate machine learning algorithms at a minimal computational cost. A new research frontier has emerged over the past decade in seismology, driven by the availability of big datasets. An abundance of machine learning algorithms has been developed for many different tasks, ranging from phase picking to signal characterization, with applications spanning diverse regions and tectonic environments. In this study, we focus on the P-phase onset characterization aiming to develop an automatic machine-learning-based workflow for focal mechanism computation that can be extended to hybrid moment tensors based on P-phase onsets.   To accomplish this, we train and assess a convolutional neural network designed to compute first-motion polarities utilizing the seismic waveform archive of the SED. We compare and evaluate three models: i) a model trained with a large first motion data set from Southern California (SoCal); ii) a model trained with only Swiss first motion data, i.e., with a significantly smaller but local data set; iii) the SoCal model, fine-tuned with the Swiss data. We compare the predicted polarities from all three models with manually picked polarities for earthquakes recorded since 2009 in Switzerland. The results obtained with the SoCal model show high precision (>90%), but only moderate recall (80%), indicating the necessity of re-training the model. To facilitate this, and considering transfer learning, we compile an augmented dataset exclusively comprising first motions from manually reviewed focal mechanisms up to 6.2023. The new model is tested on first motion data collected post 6.2023 to validate its performance. Subsequently, the model is applied to the entire manually picked Swiss dataset. Preliminary findings suggest a potential doubling of the focal mechanisms catalog, promising insights into the orientation of previously unresolved small structures. Our goal is to implement the new model in near real time, particularly in regions with dense seismic networks and high seismicity rates. This aims towards achieving near-automatic focal mechanism computation and monitoring changes in fault orientation during energetic seismic sequences.