Professional Documents
Culture Documents
Abstract
The Bayesian earthquake analysis tool (BEAT) is an open-source Python software to con-
duct source-parameter estimation studies for crustal deformation events, such as earth-
quakes and magma intrusions, by employing a Bayesian framework with a flexible
problem definition. The software features functionality to calculate Green’s functions
for a homogeneous or a layered elastic half-space. Furthermore, algorithm(s) that
explore the solution space may be selected from a suite of implemented samplers. If
desired, BEAT’s modular architecture allows for easy implementation of additional fea-
tures, for example, alternative sampling algorithms. We demonstrate the functionality
and performance of the package using five earthquake source estimation examples:
a full moment-tensor estimation; a double-couple moment-tensor estimation; an esti-
mation for a rectangular finite source; a static finite-fault estimation with variable slip;
Cite this article as Vasyura-Bathke, H.,
and a full kinematic finite-fault estimation with variable hypocenter location, rupture
J. Dettmer, A. Steinberg, S. Heimann,
velocity, and rupture duration. This software integrates many aspects of source studies M. P. Isken, O. Zielke, P. M. Mai,
and provides an extensive framework for joint use of geodetic and seismic data for H. Sudhaus, and S. Jónsson (2020). The
Bayesian Earthquake Analysis Tool,
nonlinear source- and noise-covariance estimation within layered elastic half-spaces. Seismol. Res. Lett. XX, 1–16, doi: 10.1785/
Furthermore, the software also provides an open platform for further methodological 0220190075.
development and for reproducible source studies in the geophysical community. Supplemental Material
Source Time GF
Name UQ, Samplers Source Types Function (STF) Data Types Support References
MTfit Bayesian, MH, Single source, linear full MT, No Wave polarity, No Pugh and
RJMCMC, and DC wave amplitude White (2018)
and MC ratios
GBIS Bayesian, AMH Multiple sources, isotropic point, No Static displacements Half-space Bagnardi and
spherical, rectangular, penny- (GNSS, InSAR) only Hooper (2018)
shaped crack, and ellipsoid
Bayes ISOLA Bayesian, grid Single source, quasi-nonlinear No Seismogram Yes Vackář et al.
search, and LSQ full MT waveforms (2017)
BEAT Bayesian, AMH, Multiple sources, nonlinear full MT, Half-sinusoidal, Seismogram Yes This work
SMC, and PT rectangular source, isotropic MT, triangular, and waveforms,
CLVD, DC, and full static or boxcar static displacements
kinematic finite fault (GNSS, InSAR)
AMH, adaptive Metropolis–Hastings; Bayes ISOLA, Bayesian isolated asperities; BEAT, Bayesian earthquake analysis tool; CLVD, compensated linear vector dipole; DC, double
couple; GBIS, geodetic Bayesian inversion software; GF, Green’s function; GNSS, Global Navigation Satellite System; InSAR, Interferometric Synthetic Aperture Radar, LSQ, least-
squares inversion; MC, Monte Carlo random sampling; MH, Metropolis–Hastings; MT, moment tensor; MTfit, moment tensor fit; PT, parallel tempering; RJMCMC, reversible
jump Markov chain Monte Carlo; SMC, sequential Monte Carlo; UQ, uncertainty quantification.
representations that have been used is large, even for the same of parameters (Monelli and Mai, 2008; Razafindrakoto and Mai,
earthquake, as illustrated in the source models database (Mai 2014) and prior information can be consistently integrated in
and Thingbaijam, 2014). the problem definition (Xu et al., 2015; Dutta et al., 2018).
For meaningful interpretation of estimation results, para- Only a few software packages are available for nonlinear
meter uncertainties should be quantified. These uncertainties Bayesian inference combining static and dynamic displace-
are mostly caused by measurement errors and theory errors ment data (Table 1). For example, moment tensor fit (MTfit)
(Tarantola and Valette, 1982). The measurement error is partly samples the parameters of a full MT while fixing the centroid
due to data being contaminated by random noise during meas- location (Pugh and White, 2018). The algorithm considers
urement. This noise at seismic stations can be reduced, for amplitude ratios and the polarity of seismic phases. Although
example, by employing better instruments and recording in it does not support seismic waveforms, the software is useful
low-noise environments. The most important error compo- for seismic events of low-moment magnitudes (M w ∼ 3).
nents in InSAR data are of atmospheric origin and may be Bayesian isolated asperities (BayesISOLA) is a program that
reduced, for example, through using additional independent supports seismic waveforms to sample the parameters of a full
data or through advanced processing strategies (Bekaert et al., MT in a mixed linear–nonlinear setting (Vackář et al., 2017).
2015). Theory errors arise from assumptions in the mathemati- The location and source time are sampled via grid search and
cal formulation relating the model parameters to the observed the moment tensor components are determined with a linear
data. For example, layered models that simplify the velocity least-squares inversion. Geodetic Bayesian inversion software
structure of the Earth are used in the computation of Green’s (GBIS) provides a Bayesian framework for sources with simple
functions (GFs, e.g., Yagi and Fukahata, 2011; Minson et al., geometrical shapes (prolate spheroid, penny-shaped crack,
2013; Duputel et al., 2014). Also, the parameterization of the etc.) (Bagnardi and Hooper, 2018) that have analytic closed
deformation source, like the discretization of the earthquake form solutions, assuming a homogeneous elastic half-space
fault plane into a fixed number of patches, introduces theory Earth structure. Because of these source parameterizations,
errors (Dettmer et al., 2014). GBIS is limited to static surface displacement data, but can
Parameterizations involving the source location, geometry, be useful in volcano geodesy (Table 1). Although the advan-
and source time function are nonlinearly related to the surface tages of using Bayesian inference in deformation source esti-
displacements caused by the source. Consequently, inferring mations are undeniable, a software package that is sufficiently
the source parameters and uncertainties from displacements general for a wide range of applications has been lacking.
requires nonlinear evaluation methods, which have high- In this work, we present the Bayesian earthquake analysis
computational costs (Dettmer et al., 2007). Bayesian inference tool (BEAT), a general suite of programs for a wide range of
is increasingly being applied to solve nonlinear estimation prob- linear and nonlinear deformation source studies (Tables 1 and
lems because it has the advantage of quantifying uncertainties 2). BEAT provides the first open-source implementation for
Full MT I mnn , mee , mdd , mne , mnd , med , and moment magnitude Stähler and Sigloch (2014)
DC Strike, dip, rake, and moment magnitude Jost and Herrmann (1989)
CLVD Dip, azimuth (of largest dipole), and moment magnitude Jost and Herrmann (1989)
Rectangular source Strike, dip, rake, length, width, slip, optional: Haskell (1964)
hypocentral location (x, y), and rupture velocity
CLVD, compensated linear vector dipole; DC, double couple; MT, moment tensor.
an appropriate source–receiver volume, which requires that the of the same type, each with a full and independent set of source
source region is discretized with a GF grid. model parameters, allowing for custom and complex model
The GF databases are setup such that different programs setups. With this flexible setup, complex source shapes such
can use them. Currently, support is included for quick seismo- as caldera ring faults (e.g., Bathke et al., 2015), faults with var-
grams (QSEIS) (Wang, 1999), quick seismograms for spherical iable strike and dip (e.g., R. Dutta et al., unpublished manu-
earth models (QSSP) (Wang et al., 2017), and post-seismic script, 2019, see Data and Resources), and curved dike
Green’s functions and post-seismic deformation computation intrusions (e.g., Chadwick et al., 2011) can be studied.
(PSGRN/PSCMP) (Wang et al., 2006). QSEIS calculates seis-
mograms based on a layered viscoelastic half-space model and Estimation of the slip history on a finite fault
should be used if local or regional setups are of interest. QSSP For a finite-fault source, under the assumption that the source
calculates seismograms based on a layered, self-gravitating fault geometry is known, the aim is to resolve the spatiotem-
Earth and should be used for teleseismic setups or if interaction poral slip pattern on subfaults. These sources are commonly
of the crust and atmosphere is important. PSGRN/PSCMP cal- discretized more finely than can be resolved by data (overpar-
culates synthetic stress, strain, displacement, tilt, and gravita- ameterization), which requires regularization to avoid unstable
tional fields on a layered viscoelastic half-space and should be solutions due to the ill-posed inverse problem (Altman and
used for static displacement data. Gondzio, 1999). Regularization is typically applied via smooth-
Therefore, this infrastructure allows computing geodetic ness constraints based on a trade-off between slip smoothness
static displacements (at any depth) and seismic waveforms at and data fit (Jónsson et al., 2002). In Bayesian inference,
any distance and depth range with desired frequency content. smoothing regularization can be achieved via a Gaussian prior
In addition, the open-source approach of BEAT makes exten- psjα, which includes a covariance matrix with off-diagonal
sions through user contributions straightforward to include. terms, whose Cholesky decomposition is equivalent to a
Laplacian finite-difference operator L of size M × M. The
Estimation of source location and geometry trade-off parameter α scales the degree of smoothing and can
The first-order parameters of a deformation source are source be assumed to be an unknown parameter (i.e., a hierarchical
location (east, north, and depth), source time function (time scaling that is estimated based on data information). Including
and duration), and source geometry. These are determined in smoothing regularization of slip values s on M discretized fault
a nonlinear estimation (equation 5), and the number of source patches, the PPD becomes (Fukuda and Johnson, 2008)
parameters m depends on the choice of deformation source.
BEAT supports a variety of point sources: a full MT, an iso- EQ-TARGET;temp:intralink-;df6;320;184 pm; σ 1 ; σ 2 ; …; σ K jdobs;1 ; dobs;2 ; …; dobs;K
tropic MT, a double-couple MT, and a compensated linear vec- Y
K
tor dipole (CLVD) MT (Table 2). By superposition of MTs, the ∝ pm × psjα × pdobs;k jm; σ k ; 6
GF database infrastructure permits discretizing finite sources k1
of arbitrary geometry and temporal evolution. For example, the
parameters of a rectangular source (Haskell, 1964; Heimann, with
2011) with unknown rupture nucleation and rupture velocity
can be inferred (Table 2, Fig. S1, available in the supplemental jLT Lj1=2 1
psjα exp − LsT
Ls : 7
2πα2 M=2 2α2
EQ-TARGET;temp:intralink-;df7;320;97
Example 1: full MT
Here, we used the geometry mode to estimate parameters of
an M w 5.5 MT source from simulated seismic data at regional
distances up to 1000 km (Fig. S2). Simulated waveforms were
computed using a double-couple MT source with a sinusoidal
source time function (on 0; π), and the simulated waveforms
(Figs. S3 and S4) were contaminated with simulated, tempo-
rally, and spatially correlated noise. To obtain such correlated
noise for each waveform, we summed up 300 synthesized
waveforms containing the signals of random, full MT sources
in the epicentral region (3 km) with magnitudes between 3.0
Figure 2. An example of a directory tree for a BEAT project after all and 4.7, and including five larger magnitude sources between
steps have been executed. The config_*.yaml files are the 4.5 and 5.0 with random source time variations between −20
parameter files to configure and customize the sampling process.
and 5 s with respect to the reference event. The corresponding
The results directory (created by subcommand export)
contains sampling results, estimated covariance matrices, and signal-to-noise ratio varies between 1.2 and 3.3. We applied
synthetics. The hypers directory contains sampling results of the SMC sampling with an MTSource parameterization to sam-
initial hyperparameter estimation. The linear_gfs directory ple the parameters of a full MT, its centroid location, and its
contains the Green’s functions (GFs) for the finite-fault inversion source time function.
(created by command build_gfs). stage_* folders contain To demonstrate the influence of different parameteriza-
results of the different sampling stages of the sequential Monte
tions for the data error covariance matrix (see the Seismic
Carlo (SMC) sampler (* represents the integer number of the
corresponding stage) and stage_−1 contains the PPD of the Covariance Matrix Structure section), we compared results
model parameters of the corresponding final stage. of sampling based on a diagonal covariance matrix (uncorre-
lated noise, variance type in BEAT) and a non-Toeplitz, full
covariance matrix, which accounts for time-dependent and
nonstationary noise (Dettmer et al., 2007). The comparison
during the course of the sampling and the figures directory shows that the double-couple mechanism was not retrieved in
after the results are plotted. the estimation when neglecting noise correlations (Fig. 3a,c),
but was well recovered when using a non-toeplitz covari-
Application Examples ance matrix (Fig. 3b,d). With only data noise contamination in
To demonstrate the capabilities of BEAT, we present five different this synthetic test, the uncertainty of the solution is low for the
source estimation examples. For the first example, we generated variance covariance matrix case and the data are overfitted
seismic waveforms from an MT point source and contaminated (Figs. S5 and S6). However, this overfitting was reduced using
them with synthetic correlated noise. These synthetic data were non-Toeplitz covariance matrices, and the resulting parameter
then used to estimate the components of a full MT. In the other marginals are less biased (Figs. S7–S9) than when diagonal
four examples, we considered both geodetic and seismic data of covariance matrices were used.
the 2009 L'Aquila earthquake. First, we estimate the location For this example, we used four CPU cores on a standard
and the parameters of a double-couple point source. Then, we mobile computer. We configured the SMC sampler with 2000
employed a rectangular source to estimate the fault geometry Markov chains (MCs) over 300 steps over 40 stages. During the
of the rupture assuming uniform slip. With the MAP parameters sampling, the forward model was evaluated 24 million times,
from the rectangular fault solution we then estimate spatially which took about 4 hr.
variable static slip, while keeping the fault geometry fixed. Finally,
we estimated kinematic parameters of the rupture based on the Example 2: double-couple MT
results of the previous two steps. Here, hypocentral location, rup- We considered data from the 2009 L'Aquila earthquake to esti-
ture velocity, and rupture duration were treated as unknowns. mate the double-couple moment tensor (DCSource) using
These scenarios are fully reproducible and available as step- the geometry mode of BEAT. We included P waveforms
by-step tutorials (see Data and Resources). from 35 seismic stations at teleseismic distances between
The uncertainty quantification capability of BEAT, and its 30° and 90° (Fig. S10). The seismic data were filtered with a
flexibility can result in significant computational costs. The Butterworth band-pass filter between 0.01 and 0.2 Hz. We
overall computation time depends on the amount of data, the applied a time window of −10 to 40 s around the theoretical
(b)
were used in the estimation. Patches with high-slip values have Figure 8. Results of example 5. Geodetic InSAR data fits for the
shorter rupture durations (e.g., patch 51) with the 95% confi- 2009 L'Aquila earthquake from (a) descending and (b) ascending
dence bounds between 0.1 and 2 s. The rupture velocities of acquisition geometry. (Data panels) Geocoded unwrapped
interferograms in radar line of sight (LoS) with negative values
patches close to the hypocenter (e.g., patch 67) are slightly bet-
indicating increasing LoS distance due to subsidence. The dis-
ter constrained (with the 95% confidence bounds between 2.45 played displacement values are derived from quadtree subsam-
and 3:8 km=s) than patches further away. This kinematic sol- pling and extrapolated to each pixel that belongs to the same
ution includes an estimation of the smoothing weight α, and quadtree square. The look-vector and heading of the satellite are
the result is based on a range of α values rather than a single- shown by the two short and long arrows, respectively. (Model
panels) Synthetic surface displacements in LoS derived from the
fixed smoothing value (Fig. 7c). The moment rate function
MAP solution. The small gray rectangle shows the location and
shows a source rupture duration of ∼9–10 s with the peak orientation of the derived fault geometry from example 2,
moment release occurring between 4 and 6 s (Fig. 7b). The whereas the red rectangle shows the extended fault geometry
fit with the geodetic data is comparable to example 4 (Fig. 8), that is used in examples 4 and 5. The solid black lines mark the
but the fit with the seismic waveforms (Fig. 9) appears to be upper edge of the fault. (Residual panels) Residual surface dis-
poorer than in examples 2 and 3. The latter comparison, how- placements, that is, the difference between the model and the
data panels. Note the different color scale for the residual plots.
ever, is unfair as the seismic data in this example include
significantly more details due to the increased upper cutoff fre-
quency of the band-pass filter.
Several features from our results generally agree with pre- indicates a slight rotation of the slip vector with increasing
vious studies. The rupture exhibits predominantly a single-slip depth (Zhang et al., 2012).
region at depths between 5 and 16 km. The hypocenter is In this example, we employed 25 CPU cores on a worksta-
located in the northwest and offset from the main slip region tion to sample 10,000 MCs for 300 steps over 35 stages of the
(Cirella et al., 2009; Zhang et al., 2012). The moment rate func- SMC sampler. This resulted in ∼100 million forward model
tion (Fig. 7) shows a small peak near 2 s. Finally, the rake angle evaluations and a runtime of ∼85 hr.