Understanding deep crustal structure can provide us with insights into tectonic processes and how they affect the geological record. The deep crustal structure can be studied using a variety of seismological techniques such as receiver function analysis, and surface and body wave tomography. Using models of crustal structure derived from these methods, it is possible to delineate tectonic boundaries and regions that have been affected by similar processes. However, often velocity models are grouped in a somewhat subjective manner, potentially meaning that some geological insight may be missed. Cluster analysis, based on unsupervised machine learning, can be used to more objectively group together similar velocity profiles and, thus, put additional constraints on the deep crustal structure. In this study, we apply hierarchical agglomerative clustering to the shear wave velocity profiles obtained by Gilligan et. al. (2016) from the joint inversion of receiver functions and surface wave dispersion data at 59 sites surrounding Hudson Bay. This location provides an ideal natural laboratory to study Precambrian tectonic processes, including the 1.8Ga Trans-Hudson Orogen. We use Ward linkage to define the distance between clusters, as it gives the most physically realistic results, and after testing the number of clusters from 2 to 10, we find there are 5 main stable clusters of velocity models. We then compare our results with different inversion parameters, clustering schemes (K-means and GMM), as well as results obtained for profiles from receiver functions in different azimuths and found that, overall, the clustering results are consistent. The clusters that form correlate well with the surface geology, crustal thickness, regional tectonics, and previous geophysical studies concentrated on specific regions. The profiles in the Archean domains (Rae, Hearne, and Superior) were clearly distinguished from the profiles in the Proterozoic domains (Southern Baffin Island and Ungava Peninsula). Further, the crust of Melville Peninsula is found to be in the same cluster as the crust of the western coast of Ungava Peninsula, suggesting a similar crustal structure. Our study shows the promising use of unsupervised machine learning in interpreting deep crustal structures to gain new geological insights.

Indian plate passed over four plumes during its way towards Eurasian plate. These interactions with plumes certainly affected the upper mantle and crust of Indian plate as manifested by its thinned lithosphere and relatively low shear velocities both in crust and upper mantle, but the depth extent of these effects of plume-lithosphere interaction remains ambiguous. In this study, we investigate the mantle transition zone beneath Indian shield using P-to-S receiver functions computed at 24 stations covering the entirety of the Indian shield to investigate the depth extent of the imprints of the plume-lithosphere interaction as well as to study the lateral variations of transition zone beneath a stable intraplate setting like the Indian shield. Our results show good agreement with the results of previous studies as well as with the tomographic models in terms of the average apparent depths of the 410 and 660 discontinuities and the transition zone thickness. However, unlike previous studies, we find a compelling evidence of a persistent mid transition zone discontinuity beneath all the stations and a low velocity layer beneath some regions. We also investigated the frequency dependence of amplitudes of receiver functions and found most of the stations showed strong dependence of amplitudes on frequency. Based on the evidence from our investigation, we demarcate the regions potentially containing relatively more weight percentage of water inside the otherwise considered ‘dry’ mantle transition zone. These regions should be further investigated in detail by a dense seismic network and a realistic 3-D velocity model.

For estimation of surface wave group velocity at a given period (T), the epicentral distance is divided by the difference in the arrival time of the corresponding group and the origin time of the earthquake. Hence, it is assumed that such waves are generated at the source/epicenter of the earthquake. However, this assumption is not correct. This work describes an effort to understand and quantify the amount of error that can creep into the estimated surface-wave group velocity values due to wrong assumptions used during their estimation. The error may affect the group velocity values especially when regional earthquake data is used for velocity estimation. The analysis was carried out using a horizontal layer over the halfspace model. Errors were estimated at different epicentral distances and periods for different layer thicknesses (H). Also, focal depth (h) was varied from 5 km to just 5 km above the layer boundary for each model. It is observed that for any combination of h, H, and T, error in estimated group velocity decrease rapidly with epicentral distance. The present work gives us some idea about what is the minimum epicentral distance from which data can be included for estimation of group velocity without adding significant error with such a wrong assumption. It is observed that the minimum epicentral distance at which error becomes less than or equal to a given percentage error decreases with increasing focal depth, i.e. lower is the difference between crustal thickness and focal depth, the lower is the error at a given epicentral distance and period. This means that when the difference between crustal thickness and the focal depth is low, even local earthquake data may be used without adding much error in the estimated group velocity values. For a given value of crustal thickness, focal depth, and epicentral distance, error increase with increasing period.

Seismic event classification can be challenging in the regions where different types of seismicity overlap in space, time, and magnitude. In this paper, I evaluate the performance of a supervised machine learning technique called Random Forest for the discrimination of icequakes and earthquakes in southeast Alaska at 15 stations surrounding the region. I train the Random Forest on about 3000 icequakes and earthquakes that occurred in the region over the last 17 years. For each event, absolute frequency spectrum values are considered as input features. The accuracies at different stations range from 75 to 95% with an average of about 90%. I conducted tests for selecting the optimum number of decision trees in the RF model and compared the results obtained by applying bandpass filters of different frequency bands on input waveforms. I further experiment by reducing the dimensions of input features by applying Principal Component Analysis (PCA), and conducted test for selecting the minimum number of components and the frequency band that gives the best results. The application of PCA resulted in slightly better results and a final model that gave the best results among all the tests was chosen. The accuracy results of the final model were further analyzed with respect to the amount of available dataset, the average distance of a station from all the glaciers, and the local geology.