Application of a machine-learning model for the determination of focal mechanisms in the area of the Corinth Rift Laboratory Near-Fault Observatory (CRL NFO), Central Greece

Research Highlights
•     First motion polarities determined using an ML model yielded results comparable with manual measurements.
•     Discrepancies were mainly observed at stations in larger epicentral distances or equipped with accelerometers.

Introduction
The large quantities of available seismic waveform data have rendered their manual analysis a challenging task. The
unification of seismic networks in Greece (e.g., Evangelidis et al., 2021) has improved the completeness of the routine
analysis seismic catalogs. However, as the number of earthquakes is multiplied almost tenfold for every lowering of a
magnitude unit, manual analysis can only go so far before the workload becomes prohibiting. Conventional automatic
seismic P- and S-wave arrival-time picking methods also have downsides, depending upon the objective function
used in each case. Despite providing quick solutions, the measurements of these automatic methods are usually
less accurate than those performed by human analysts, and they need to be manually revised for their quality to
be improved. However, in recent years, the rise of Machine-Learning (ML) models has revolutionized the automatic
processing of seismological data. ML models, such as PhaseNet (Zhu & Beroza, 2019) and EQTransformer (Mousavi
et al., 2020), are now used for the automatic analysis of huge datasets, greatly augmenting the number of detected
events and lowering the completeness magnitude of seismic catalogs. Even more surprisingly, in terms of accuracy,
these ML methods provide results that are on par with manually analyzed data (e.g., Mousavi et al., 2020) and can
even recognize P- and S-wave arrivals in noisy signals without filter application (e.g., Zhu & Beroza, 2019).
An even more challenging task is the automatic identification of P-wave First Motion Polarities (FMPs). For earthquakes
of moderate or higher magnitude (e.g., M33.6), the focal mechanisms can usually be determined through moment
tensor inversion, by fitting synthetic to the observed waveforms. FMPs, on the other hand, are useful for determining
focal mechanisms for weak events (M£3.5), particularly when the network density is adequate near the epicentral
area. The distribution of FMPs depends on the radiation pattern of seismic energy at the source. In the double-couple
model, FMPs are distributed in four quadrants, characterized by the alternating sign of the first motion recorded at
the vertical component of the seismic stations. These first pulses can be impulsive or emergent, depending on the
angle of emergence of the seismic ray at the focus, relative to the radiation pattern of the faulting, which defines the
focal mechanism; i.e., the strike, dip, and rake of the rupture. Rays originating near the P or T principal axes of the
moment tensor are expected to produce impulsive negative or positive first motions, respectively. On the other hand,
rays originating near the nodal planes or the null axis are likely to produce emergent first motions, which cannot be
easily distinguished (small, nearly zero amplitude). Furthermore, the Signal-to-Noise Ratio (SNR) of weak-magnitude
events can be too low for even a P-wave arrival to be picked. The application of filters to lower the noise level and
improve SNR can distort the first pulse and lead to an erroneous FMP measurement. Even when a P-wave arrival-
time can be roughly assessed, the analyst must decide whether to characterize the polarity of the first motion or skip
it to avoid an error. Other measurements that can help constraining the focal mechanisms of weak events include the
S-to-P amplitude ratio and the S-wave polarization direction (e.g., Kapetanidis et al., 2015).
As the characterization of the FMP is a pattern recognition task, it has also been the subject of studies employing
methods based on artificial intelligence. Such ML models can recognize these patterns and classify the first pulse
as being positive or negative, assess whether it is impulsive or emergent, or decide to dismiss a measurement if the
polarity cannot be safely determined. Examples of such ML models include those presented by Ross et al. (2018)
and Zhao et al. (2023). Both models were trained on a large dataset of seismic waveforms from Southern California,
and the authors reported a precision of ~95% or more compared to the respective FMP measurements by human
analysts.