Sequencing Seismic Data and Models

Date: 4/24/2019

Time: 11:45 AM

Room: Elliott Bay

Long-term operations of high-quality networks, completion of the EarthScope project, continuing temporary deployments, and dense arrays enabled by new technology, have yielded massive datasets of seismic waveforms ripe for analysis and interpretation. In tandem, tomographic models across scales have proliferated, as have structural constraints from scattered and converted waves and ambient noise correlations. Gleaning insights into Earth structures and processes from these vast and heterogeneous datasets requires approaches flexible enough to detect with little-to-no user supervision robust patterns within and across different types of measurements made under various observing conditions. Here, we present results from a completely new sequencing-based unsupervised approach for identifying patterns and trends in seismic models and data. The algorithm arranges the data along a 1D manifold to form a sequence that minimizes differences between neighboring data in the sequence as well as globally across the entire sequence. A datum’s position in the sequence can be used to visualize geographic patterns in the dataset – as we illustrate using group velocity dispersion curves and velocity profiles drawn from topographic models – and to identify anomalous signals and place them in context – as we do using waveforms of diffracted waves. We present results of sequencing of Love and Rayleigh dispersion curves across the conterminous United States, which show that the algorithm is able to map out large- and small-scale geographic patterns corresponding to structural boundaries within the crust and to physiographic provinces. Finally, we discuss the utility of sequencing and its advantages over more traditional techniques for exploring large datasets and identifying patterns within them.

Presenting Author: Vedran Lekic

Authors