Back to research

Earthquake catalog development

Locating Himalayan Microseismicity

An earthquake-catalog workflow combining machine learning–based phase detection with association and relocation of small Himalayan earthquakes.

  • Machine Learning
  • Signal Analysis
  • Earthquake Location

Research question

How can machine learning help extract small-earthquake signals from continuous seismic data and turn detections into physically interpretable event locations?

Study design

I developed an earthquake-catalog workflow that connects machine learning–based phase picking with phase association, absolute location, and double-difference relocation. The dense seismic array near Melamchi provides observations within the central Nepal Himalaya. I use these records to turn candidate arrivals into an earthquake catalog and examine the distribution of seismicity in its regional tectonic setting.

Four-panel tectonic context: the India–Eurasia collision, central Nepal and the 2015 Gorkha earthquake, the nodal array near Melamchi, and published Main Himalayan Thrust geometries.
The array connects local earthquake observations to the Himalayan tectonic setting. The panels move from the India–Eurasia collision to central Nepal, the Melamchi-area deployment, and published interpretations of Main Himalayan Thrust geometry. Panel (d) is adapted from Thapa et al. (2023), with the original references retained in the figure.

Workflow highlights

  • Machine learning turns continuous records into phase picks

    PhaseNet identifies candidate seismic arrivals, and REAL associates picks across stations into events. The workflow also includes QuakeMigrate as an alternative detection and location approach.

  • Velocity models connect arrival times to earthquake locations

    The location workflow includes VELEST, Hypoinverse, and a combined approach. VELEST also supports one-dimensional velocity-model inversion, including models with and without a low-velocity zone.

  • Relative locations reveal the catalog’s spatial structure

    Double-difference relocation with hypoDD refines the relative positions of nearby earthquakes. Map, depth, and three-dimensional views then bring the resulting spatial distribution into focus.

Earthquake locations in the Nepal Himalaya in four panels: a longitude–latitude map, depth profiles against latitude and longitude, and a three-dimensional view. Points are colored by local magnitude (ML).
Earthquake locations viewed from the surface, in depth profiles, and in three dimensions. Together, these views show the catalog’s horizontal and vertical distribution; color indicates local magnitude (ML).

Detailed methods

Catalog workflow: ObsPy preprocessing; PhaseNet with REAL or QuakeMigrate; VELEST, Hypoinverse, or a combined absolute-location route; then hypoDD relative relocation.
From continuous data to a relocated catalog. The workflow places machine-learning phase picking within a sequence of association, velocity-model and absolute-location calculations, and double-difference relocation. Parallel branches show the alternative approaches included at each stage.

1. Prepare continuous waveform data

I preprocess the continuously recorded seismic data with ObsPy to provide the waveform inputs for phase detection and event association.

2. Detect phases and associate events

The machine-learning branch uses PhaseNet for seismic phase picking and REAL to group arrivals across stations. QuakeMigrate provides an alternative branch of the workflow. These steps connect individual waveform detections to candidate earthquakes.

3. Estimate absolute locations and velocity structure

The workflow includes VELEST, Hypoinverse, and a combined VELEST–Hypoinverse route. In the VELEST branch, earthquake location is paired with one-dimensional velocity inversion, with options that include or omit a low-velocity zone.

4. Refine relative locations

The absolute-location branches feed into hypoDD for double-difference relocation. Relative arrival-time information refines the spatial relationships among nearby events, producing the catalog shown in the map and depth projections.

Why it matters

This project demonstrates how I integrate machine-learning tools into a scientific analysis: converting automated detections into a structured data product that can support questions about the distribution of seismicity.