A document by Pragnath Dodda. Click on the document to view its contents.

Teleseismic shear-wave splitting analyses are typically performed by reversing the splitting process through the application of frequency- or time-domain operations minimizing transverse-component waveforms. These operations yield two splitting parameters, φ (fast-axis orientation) and δt (delay time). In this study, we investigate the applicability of a recurrent neural network, SWSNet, for determining the splitting parameters from pre-selected waveform windows. Due to the scarcity of sufficiently labelled real waveform data, we generate our own synthetic training dataset. The model is capable of determining φ and δt with a root mean squared error (RMSE) of 9.58◦ and 0.143 s for noisy synthetic test data. The application to real data involves a deconvolution step to homogenize the waveforms. When applied to data from the USArray dataset, the results exhibit similar patterns to those found in previous studies with mean absolute differences of 11.12◦ and 0.25 s in the calculation of φ and δt, respectively.

The detection and rapid characterisation of earthquake parameters such as magnitude are important in real time seismological applications such as Earthquake Monitoring and Earthquake Early Warning (EEW). Traditional methods, aside from requiring extensive human involvement can be sensitive to signal-to-noise ratio leading to false/missed alarms depending on the threshold. We here propose a multi-tasking deep learning model – the Convolutional Recurrent model for Earthquake Identification and Magnitude Estimation (CREIME) that: (i) detects the earthquake signal from background seismic noise, (ii) determines the first P-wave arrival time and (iii) estimates the magnitude using the raw 3-component waveforms from a single station as model input. Considering, that speed is essential in EEW, we use up to two seconds of P-wave information which, to the best of our knowledge, is a significantly smaller data window compared to the previous studies. To examine the robustness of CREIME we test it on two independent datasets and find that it achieves an average accuracy of 98\% for event-vs-noise discrimination and can estimate first P-arrival time and local magnitude with average root mean squared errors of 0.13 seconds and 0.65 units, respectively. We compare CREIME with traditional methods such as short-term-average/ long-term-average (STA/LTA) and show that CREIME has superior performance, for example, the accuracy for signal and noise discrimination is higher by 4.5\% and 11.5\% respectively for the two datasets. We also compare the architecture of CREIME with the architectures of other baseline models, trained on the same data, and show that CREIME outperforms the baseline models.

The polarity of first P-wave arrivals plays a significant role in the effective determination of focal mechanisms specially for smaller earthquakes. Manual estimation of polarities is not only time-consuming but also prone to human errors. This warrants a need for an automated algorithm for first motion polarity determination. We present a deep learning model - PolarCAP that uses an autoencoder architecture to identify first-motion polarities of earthquake waveforms. PolarCAP is trained in a supervised fashion using more than 130,000 labelled traces from the Italian seismic dataset (INSTANCE) and is cross-validated on $\sim$22,000 traces to choose the most optimal set of hyperparameters. We obtain an accuracy of 0.98 on a completely unseen test dataset of almost 33,000 traces. Furthermore, We check the model generalizability by testing it on the datasets provided by previous works and show that our model achieves a higher recall on both positive and negative polarities.