Benchmark
Computation Performance
Qseek is built for searching in large-N data sets. The implementation is leveraging Python asyncio heavily to implement threading and keeping the CPU busy. It is built on top of highly performant Pyrocko functions implemented in C language. The inference is using PyTorch which enables GPU computation of the seismic imaging functions.
This enables high throughput of seismic data in different scenarios.
| Number Stations | Throughput in MB | Throughput in Waveform data |
|---|---|---|
| 300+ | 50 MB/sec | 12 hours/sec |
| 50 | 200 MB/sec | 6 hours/sec |
Scanning a 600 GB (~700 years of waveforms) data set costs ~2 days on a 64 cores machine equipped with an Nvidia A100 GPU.
Note
The performance depends heavily on the octree resolution and the number of events detected in the data set.
Related Projects
A list of other projects using stacking and migration approach to back-project seismic energy sources in 3D space:
Lassie-v1
Lassie - The friendly Earthquake detector in version 1. Qseek utilizes the same optimized heavy-duty functions for stacking and migration as Lassie v1.
QuakeMigrate
QuakeMigrate uses a waveform migration and stacking algorithm to search for coherent seismic phase arrivals across a network of instruments. It produces—from raw data—catalogues of earthquakes with locations, origin times, phase arrival picks, and local magnitude estimates, as well as rigorous estimates of the associated uncertainties.
BPMF
Complete framework for earthquake detection and location: Backprojection and matched-filtering (BPMF), with methods for automatic picking, relocation and efficient waveform stacking.
Loki
LOKI (LOcation of seismic events through traveltime staKIng) is a code that performs earthquake detection and location using waveform coherence analysis (waveform stacking).
MALMI
MALMI (MAchine Learning aided earthquake MIgration location), variant of Loki for detecting and locating earthquakes using ML image functions provided by SeisBench.