Three-Dimensional Modeling of the Xichang Crust in Sichuan, China by Machine Learning

Seismicity and distribution of earthquakes can provide active fault structural information on the crust at a regional scale. The morphology of faults can be derived from the epicentral distribution of micro-earthquakes. In this study, we combined both the relocated earthquake catalogue and related preliminary geophysical information for 3D modeling of the crust in the Xichang area, Sichuan province, China. The fault morphology and deep crustal structure were automatically extracted by the machine learning approach, such as the supervised classification and cluster analysis methods. This new 3D crustal model includes the seismic velocity distribution, fault planes in 3D and 3D seismicity. There are many earthquake clusters located in the folded basement and low-velocity zone. Our model revealed the topological relation between the folded basement and faults. Our work show the crustal model derived is supported by the earthquake clusters which in turn controls the morphological characteristics of the crystalline basement in this area. Our use of machine learning techniques can not only be used to predict the refined fault geometry, but also to combine the seismic velocity structure with the known geological information. This 3D crustal model can also be used for geodynamic analysis and simulation of strong motionseismic waves.

Active Volcanism Revealed from a Seismicity Conduit in the Long-resting Tatun Volcano Group of Northern Taiwan

Abundant earthquakes clustered within a particular zone often reflect an active geological feature, such as clustering seismicity along a fault zone and a huge number of volcanic-earthquakes around the erupting conduit. Herein we perform a double-difference tomographic inversion and relocate the seismicity at the long-resting Tatun volcano group (TVG) in northern Taiwan. A dramatic improvement of the earthquake location model surprisingly show that, from 2014 to 2017, two clustered seismic zones are identified in the TVG. One major group of events (>1000) persistently clustered within a ~500 m diameter vertical conduit with a ~2 km height. The clustering seismicity conduit is just located nearby Dayoukeng, one of the strongest fumaroles in the TVG, and is connected to a fracture zone characterized by low Vp/Vs in the shallow crust. The other group of events is clustered within a sphere-like zone beneath Mt. Chihsin around the depths between 0.5 km and 2 km. Both seismic zones are probably triggered by the significantly volcanic gases and fluids ascending from the deep magma reservoir. Combined with a variety of results from literature, the seismicity conduit near the strong fumarole is the evidence for an active volcano and also identifies a likely pathway for ascending magma if the TVG erupts again in the future. But possibility of developing different magma pathways at other clustered seismic zones such as beneath Mt. Chihsin may not be totally excluded.

Condensation of earthquake location distributions: Optimal spatial information encoding and application to multifractal analysis of south Californian seismicity

We present the "condensation" method that exploits the heterogeneity of the probability distribution functions (PDFs) of event locations to improve the spatial information content of seismic catalogs. As its name indicates, the condensation method reduces the size of seismic catalogs while improving the access to the spatial information content of seismic catalogs. The PDFs of events are first ranked by decreasing location errors and then successively condensed onto better located and lower variance event PDFs. The obtained condensed catalog differs from the initial catalog by attributing different weights to each event, the set of weights providing an optimal spatial representation with respect to the spatially varying location capability of the seismic network. Synthetic tests on fractal distributions perturbed with realistic location errors show that condensation improves spatial information content of the original catalog, which is quantified by the likelihood gain per event. Applied to Southern California seismicity, the new condensed catalog highlights major mapped fault traces and reveals possible additional structures while reducing the catalog length by ∼25%. The condensation method allows us to account for location error information within a point based spatial analysis. We demonstrate this by comparing the multifractal properties of the condensed catalog locations with those of the original catalog. We evidence different spatial scaling regimes characterized by distinct multifractal spectra and separated by transition scales. We interpret the upper scale as to agree with the thickness of the brittle crust, while the lower scale (2.5 km) might depend on the relocation procedure. Accounting for these new results, the epidemic type aftershock model formulation suggests that, contrary to previous studies, large earthquakes dominate the earthquake triggering process. This implies that the limited capability of detecting small magnitude events cannot be used to argue that earthquakes are unpredictable in general.