Institute of Geosciences | Geophysics | Seismology | Volcanology

Multi-Resolution seismic Attenuation

Tomography
Version 3.0

Luca De Siena
Associate Professor of Geophysics with Volcanology at JGU Mainz, Germany

Honorary Lecturer at the University of Aberdeen, UK

work phone: +49 (0) 61313928499

cell: +49 (0) 15122316065

email: [email protected]

internet: www.lucadesiena.com
MuRAT in Brief
c January 8, 2022

MuRAT is a Matlab Package for Seismic Attenuation, Scattering and Absorption Tomog-
raphy using Body and Coda Waves. The quickest way to implement MuRAT is to read
his online readme and follow step-by-step his wiki. This PDF document remains the most
complete manual for MuRAT.

Features

What it does: MuRAT is a software package for measuring seismic attenuation, scatter-
ing, and absorption from passive and active data, and model 3D variations of these
parameters in space.

The first MuRAT: MuRAT1.0 was developed by Luca De Siena during his PhD at the
INGV-Osservatorio Vesuviano, Italy, and published in 2014 while he was research as-
sistant at the Westfälisches Wilhelms Universität, Münster. Three sample papers pub-
lished using this code are: De Siena et et al. 2017, SR, Prudencio et al. 2020,
and Sketsiou et al. 2021;

In 2D: MuRAT2D is the result of the activity of the Volcano Earth Imaging group.
It produces 2D seismic scattering and absorption maps at multiple frequencies and
is suited for largely undetermined problems. Three example papers published us-
ing this code are: De Siena et et al. 2017, GRL, Napolitano et al. 2020, and
Sketsiou et al. 2020;

In 3D: MuRAT3.0 (or MuRAT3) is the current standard for measuring total attenuation,
scattering and absorption - the project started in 2021 and there are still no published
papers. This guide describes this code.

1
Contents

1 Introduction 5

2 Direct-wave attenuation: Q 10

2.1 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.2 The coda normalisation method in pills . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.3 The theory bit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.3.1 Applicability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.4 MuRAT workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3 Scattering attenuation: peak delays 17

3.1 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.2 The peak delay method in pills . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

3.3 The theory bit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3.3.1 Applicability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.4 MuRAT workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

4 Absorption: coda attenuation 24

4.1 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4.2 The Qc method in pills . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4.3 The theory bit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

4.3.1 MLTWA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

4.3.2 Sensitivity kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

4.3.3 Applicability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4.3.4 MuRAT workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2
5 Getting started 32

5.1 Installation and set up in a nutshell . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5.2 Matlab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5.2.1 Requirements and installations in pills: . . . . . . . . . . . . . . . . . . . . . 33

5.2.2 Changing folder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

5.2.3 README.md . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

5.2.4 Wiki . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5.3 Julia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

6 Data preparation 37

6.1 SAC files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

6.2 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

6.3 sac_dataf older - checking data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

6.3.1 M urat_test - testing single waveforms . . . . . . . . . . . . . . . . . . . . 39

6.3.2 M urat_testAll.m - creating a file to check all headers in dataset . . . . . . 40

6.3.3 M urat_changeHdr.m - changing headers of single files . . . . . . . . . . . 41

7 Input files 42

7.1 The fields of the input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

7.1.1 GENERAL FIELDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

7.1.2 WAVEFORM DATA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

7.1.3 PEAK DELAY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

7.1.4 DIRECT WAVE ATTENUATION . . . . . . . . . . . . . . . . . . . . . . . . 46

7.1.5 CODA ATTENUATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

7.1.6 GEOMETRY AND VELOCITY . . . . . . . . . . . . . . . . . . . . . . . . . 46

7.1.7 INVERSION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

3
8 Tests 48

8.1 Analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

8.1.1 Peak delays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

8.1.2 Coda attenuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

8.1.3 Direct-wave attenuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

8.2 Damping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

9 Rays and Kernels 55

9.1 Ray tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

9.2 Kernel implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

Appendices 59

A MuRAT input file - Mount St Helens 60

4
Chapter 1

Introduction

This document describes how to use the Matlab software package MuRAT to perform seismic tomog-
raphy: however, it does not follow the traditional path of imaging seismic velocity with travel times,
or phase/group velocities using surface-waves. MuRAT is a code for seismic attenuation tomography.

To understand the difference consider the following seismogram, produced by a volcanic earthquake:

Figure 1.1: A seismogram. MuRAT measures the heterogeneous loss of seismic energy using different
seismic attributes

We are used to picking phases on it - for example, the P-wave arrival. By modelling the source-
station path followed by the corresponding wave packet with a ray-tracing approach, we measure
velocity changes over different paths. Travel-time tomography requires a non-linear inversion, where
we need to update the paths followed by the waves, the corresponding velocity model, and seismic
locations jointly to find a reliable solution. There are multiple codes that do this: our favourites are

5
LOTOS and FMTOMO (all underlined names in the manual are links to the corresponding internet
sites). However, MuRAT is part of a wider framework working across multiple languages (Matlab,
Python and Julia). This comprises codes that allow the user to measure and image body-wave and
surface wave velocity in space (from earthquake and noise) (interfaced with python) and full waveform
inversions (in Julia). This framework especially couples seismology with outputs from Geodynamics
codes, like LaMeM, efficiently. It is highly recommended to combine MuRAT with the partner codes
for a complete imaging and modelling of the area under study.

MuRAT deals with the loss of energy suffered by seismic waves while they propagate through the
heterogeneous Earth. It fills the gap of an open-access tool to measure attenuation, separate the
processes that lead to it, and model them in space. To our knowledge, the first mention of an
attenuation tomography was given by Ho-Liu et al. (1988) in the Coso Valley. More than 30 years
afterwards, this technique is a standard with applications at the global, regional and local scales.
Yet, different mechanisms attenuate seismic waves while they propagate. While codes to model total
attenuation in space exist, no open-access code can measure attenuation, separate the mechanisms
that produce it and model their variations in space with seismic inversion. No code can certainly do
it using as input seismograms downloaded from the data servers.

We define here the mechanisms causing seismic attenuation:

1. Seismic energy progressively spreads over larger volumes while propagating far from the source.
Energy thus decreases with distance due to geometric spreading. This mechanism is independent
of frequency for a given phase. However, depending on frequency, often more than one phase
are included in a window of time measured on a seismogram - in Fig. 1.1 it is challenging to
distinguish between S-waves and surface waves produced right after them.

2. Seismic energy is lost into the heterogeneous Earth depending on the ratio between the wave-
length (λ) and the dimension of the heterogeneity (a). Energy thus decreases with distance due
to scattering losses. This mechanism depends on frequency and comprises complex physics, as
those described by multi-pathing and diffractions. Reflection and refraction can be described as
scattering processes.

3. Seismic energy interacts with the heterogeneous Earth losing heat for each cycle. Energy thus
decreases with distance due to absorption losses in the materials they pass through. This
mechanism depends on frequency. How much energy the Earth absorbs is difficult to quantify,
but such a quantification allows connecting seismology with other disciplines.

6
 4. The sum of scattering and absorption losses gives total attenuation. This is the attenuation
experienced by direct wave-packets while they propagate through the heterogeneous Earth. When
we talk about seismic attenuation we generally refer to total attenuation. However, this is only
valid in specific circumstances and for specific phases.

MuRAT measures and models total attenuation, scattering attenuation and

absorption in space using single seismogram observations. The first three
chapters of this documentation offer an overview of the theory underlying
attenuation tomography in these three forms.

In chapter 2, we revise the theory behind measurement and modelling of the energy lost by direct
waves (P and S, Fig. 1.1). The method used to model total attenuation in chapter 2 is the coda
normalisation (CN) method (Aki, 1980; Yoshimoto et al., 1993). In tomography, the method has been
introduced by Del Pezzo et al. (2006). At the core of MuRAT, there is this seminal paper, which
has been developed by De Siena et al. (2009) and De Siena et al. (2010) for muti-scale applications.
De Siena et al. (2010) benchmark the CN with the standard spectral slope (t∗ ) method, applied by
the global community. For a decade, the method has only been applied to volcanoes. The reason is
the need for a relevant coda (tc , Fig. 1.1) to normalise direct-wave information. The method has the
great advantage of removing source and site effects from direct wave measurements, but suffers from
an inexact description of the process. MuRAT uses the standard description of geometric spreading,
rα , with r is the hypocentral distance and α the geometric spreading coefficient, measuring α with its
uncertainty. By comparing the results to the assumed value of α for direct wave energies (α = 1) the
user knows how far this is from the theory. MuRAT3D deals with the problem associated with direct-
wave measurements and modelling, using theoretical and computational results obtained throughout
the last decade.

In chapter 3, we revise the theory behind the use of peak delays as markers of scattering losses (Saito
et al., 2002; Takahashi et al., 2007, 2009; Tripathi et al., 2010; Calvet and Margerin, 2013; Calvet
et al., 2013; De Siena et al., 2016; Sato, 2016). Peak-delays are a focus of the early-warning community
and a quantity important to seismic source modelling in the near field. The standard theory recognises
peak delay as a marker of scattering losses in the far field, modelled by the Markov approximation
(Saito et al., 2002), in the case scattering is produced only by long wavelength component (λ << a.
However, recent research also suggests a dependence of peak-delay on scattering when λ is of the order
of propagation distance and correlation length of heterogeneity (Napolitano et al., 2020; Di Martino

7
et al., 2021). Here, phenomena like diffraction and trapping take over. The chapter discusses the
standard application, theoretical limitations and future outlooks.

In chapter 4, we revise the theory behind the use of late-time coda attenuation as marker of seismic
absorption (Wegler and Luehr, 2001; Prudencio et al., 2013; Calvet and Margerin, 2013; Del Pezzo
et al., 2016; De Siena et al., 2016, 2017a; Del Pezzo et al., 2018; Sketsiou et al., 2020). Its application
relies on a preliminary Multiple-Lapse-Time-Window-Analysis (MLTWA - Fehler et al. (1988)).Coda
attenuation can be described using single and multiple scattering. MuRAT provides the tools to test
the onset of equipartition, a necessary condition for diffusion, where coda attenuation is equivalent
to seismic absorption. The embedded sensitivity kernels model coda attenuation in space based on
Paasschens equations (Paasschens, 1997) and the approximations of Pacheco and Snieder (2005). The
theory is inexact in the case of sharp vertical or lateral boundaries, so much so that diffusion could
never onset Morozov (2008). However, in zones where the Moho is thick and for sufficient amount
of data, the theory generally produces stable results (Sketsiou et al., 2020). New work has been
recently developed to include the effect of sharp boundaries and vertical velocity variations on coda
attenuation measurements (Sanborn et al., 2017; Cormier and Sanborn, 2019; Nardoni et al., 2021).
MuRAT currently provides the tool to invert for coda attenuation using a diffusive approximation and
sensitivity kernels.

From chapter 5, this manual describes how to install and use the code, so you
can start from there if you already know the theory.

Chapter 5 describes installation procedures and online documentation tools for MuRAT.
MuRAT 3.0 and its Readme.txt are available on GitHub and has a Wiki explaining the primary steps
c
to run the code and interpret results. We discuss why the code is released in proprietary (Matlab )
language and the corresponding installation procedures.

Chapter 6 describes data preparation for MuRAT. MuRAT 3.0 accepts only Seismic Analysis Code
(SAC) data. SAC is a standard for seismologists, and the favourite format when converting miniseeds
on data servers like IRIS. It is possible you have been processing data with another software. In
this chapter, we will discuss the conversions necessary to create a readable dataset for MuRAT and
especially the tools you can use to assess if the data are in the right format, created directly in Matlab.

Chapter 7 describes the input files for MuRAT. We offer three example applications (Mount St.
Helens, US - Toba, Indonesia - Vrancea, Romania). In theory, these are the only files the users need
to edit, once data files are in the correct format.

8
Chapter 8 describes the implementation of 3D velocity models, which is optional for MuRAT.

Chapter 9 describes the tests offered by MuRAT to check the reliability of the results.

Chapter 10 describes the regionalisation and inversion procedures.

Chapter 11 describes file outputs and plotting.

9
Chapter 2

Direct-wave attenuation: Q

Timeline in papers.

Paper Descriptor Where

(Aki and Chouet, 1975) S-wave CN method Link

(Yoshimoto et al., 1993) P-wave CN method Link

(Del Pezzo et al., 2006) S-wave CN tomography Link

(De Siena et al., 2010) Benchmark with t∗ method Link

(De Siena et al., 2014a) MuRAT1.0 for CN method published Link

(De Siena et al., 2014b) P-wave coda-normalisation tomography Link

(Prudencio et al., 2015) Active CN tomography Link

(De Siena et al., 2017b) CN tomography measuring Qc Link

10
2.1 Parameters

Input Parameters for MuRAT

Symbol Descriptor Where

tP or tS (s) P-wave or S-wave arrivals SAC Haeder

tl (s) Length of P- or S-wave window Set in input file

t0 (s) Origin time SAC Haeder

tc (s) Lapse time from origin time Set in input file

tw (s) Length of coda window Set in input file

2.2 The coda normalisation method in pills

The method measures direct and coda-wave energies and divides them to obtain a quantity that only
depends on Q, allowing for a linearized inversion.

Pros of measuring direct-wave attenuation using coda-normalised energies

1. Single-station measurements - you can get all the info from a single seismogram.

2. No need for correcting source and site effects - these are particularly relevant in the heterogeneous
Earth.

Cons of measuring direct-wave attenuation using coda-normalised energies

11
 1. Uncertain sensitivity of coda waves to Earth structures - the assumption of homogeneous coda
wave sensitivity is generally unfulfilled. Tackled by MuRAT estimating the measurable coda-wave
energy at the chosen lapse time and assuming a diffusive behaviour.

2. Correction of radiation pattern - increasing direct wave windows smooths it at the expense of
focusing. Tackled by working with different windows to measure direct-wave energy.

Figure 2.1 is a visual representation of the method.

Figure 2.1: The CN method measures the spectral energies of direct and coda waves and inverts for
Q. In red the P and S wave energies and the normalizing coda energies.

2.3 The theory bit

The coda-normalisation (CN) method normalises P- and S-wave energy for coda energy (Aki, 1980;
Yoshimoto et al., 1993). Seismic wave energies can be modelled as the convolution of the source,
path, site, and instrument functions, so that the spectral energy of direct waves (E) is:

 
−γ 2πf
E(f, r) = Rθφ SS (f )r I(f )G(f, Ψ) exp − r (2.1)
Q(f )v

12
where f is the frequency, r is the source-receiver distance, tc is the coda-wave central lapse time from
the origin time of the event (t0 ), Rθφ is the source radiation pattern (θ and φ the azimuth and take-off
angle for a source-receiver ray), SS (f ) is the source function, γ the geometrical spreading exponent,
I(f ) the known instrumental response, G(f, Ψ) is the site amplification factor (with Ψ being the
incident angle of the ray at the station), Q is the direct-wave quality factor and v the average velocity
in the medium (Yoshimoto et al., 1993).

The spectral energy of coda waves (Ec ) can be written as:

Ec (f, tc ) = SS (f )P (f, tc )G(f )I(f ). (2.2)

The attenuation of coda waves can then be expressed as (Aki and Chouet, 1975):

P (f, tc ) ' t−n −1

c exp −2πf Qc tc (2.3)

In this equation, n is the envelope spectral decay, Qc is the coda quality factor, and Ac (f, tc ) does not
include the effect of the source radiation pattern. We will discuss moire about this term in chapter 4.

Eq. 2.1 can be divided by Eq. 2.2 to normalize the energy of the spectral energy of direct waves using
the spectral energy of coda waves:

 
E(f, tc ) −γ G(f, Ψ) 2πf 1
=r exp − r (2.4)
Rθφ Ec (f, tc ) G(f ) Q(f )v P (f, tc )
where the source function SS (f ) and the instrumental response I(f ) contributions disappear. By
measuring direct-wave energy over a window of sufficient length (tl ), we smooth the azimuthal contri-
bution of the radiation pattern (De Siena et al., 2009, 2010). If the length of the direct-wave window
is chosen appropriately, the contribution of the source radiation pattern Rθφ is negligible (De Siena
G(f,Ψ)
et al., 2009) so that Rθφ = 1 and G(f ) = 1. Since early coda consists of randomly scattered waves,
the coherency and the source radiation pattern is eventually lost (Takemura et al., 2009). De Siena
et al. (2010) demonstrate that a coda window of 2 s is sufficient to make the radiation pattern quasi-
isotropic in a volcanic caldera. Still, the window length must be carefully chosen at each frequency to
avoid near-receiver onset of surface waves (Gabrielli et al., 2020).

At fixed frequency band and starting lapse-time for coda windows, P (f, tc ) is assumed to be constant
in the standard coda-normalisation method (Del Pezzo et al., 2006; Sato et al., 2012). Taking the

13
logarithm, Eq. 2.4 becomes:

 
E(f, tc ) 2πf
ln = − ln P (f, tc ) − γ ln (r) − r, (2.5)
Ec (f, tc ) Q(f )v
a linear equation solved for three unknown (ln P (f, tc ), γ, Q) once the hypocentral distance is known
and the frequency is set. If Qc is known for each source station pair, one can pre-estimate P (f, tc )
using equation 2.3, as done by De Siena et al. (2017b):

1 E(f, tc )) 1 γ ln (r)
ln + ln P (f, tc ) = − − Q(f )−1 t, (2.6)
2πf Ec (f, tc ) 2πf 2πf
where t is the travel time of the corresponding direct phase. The system can be solved with a linear
inversion for γ and Q(f )−1 . This is the method implemented in MuRAT3.0.

2.3.1 Applicability

For the applicability of the CN method, one must measure direct-wave energy. This seems simple
enough, but it can become very problematic depending on frequency, heterogeneity and source-station
distance. To check the validity, one must always look at relationship between direct-wave energy and
epicentral distance. Is there a dependence? If not, near field or strong scattering might be prevalent.
Does the relationship follow standard geometrical spreading, like those for direct wave energy? How
important is the surface wave component? A positive is that generally where you cannot measure
direct-wave energy, you can use codas, i.e., diffusion onsets pretty quickly (see chapter 4).

2.4 MuRAT workflow

MuRAT works in analogy to De Siena et al. (2010), De Siena et al. (2014a) and Sketsiou et al.
(2021) when mapping peak delays. These authors show that one must check the relationship between
hypocentral distance (or travel time) and measured energy ratio. In Figures 2.2, we show an example
at Mount St. Helens volcanoes, where sources and stations can be either on or inside the volcanic
edifice and near surroundings (red dots) or far from them (cyan dots). The expected decrease of
energy with hypocentral distance is unfulfilled at small distances, where surface and trapped waves
dramatically increase coda energy.

The energy in the coda can be back-traced to mark the position and intensity of the corresponding

14
Figure 2.2: A plot of the logarithm of the P-wave direct-to-coda energy ratios (red and cyan dots,
computed at MSH in three frequency bands) versus the P-wave travel times at Mount St. Helens
volcano. The red lines show the linear fit obtained by using only data between travel times of 0 and 2
s, affected by surface waves inside the volcanic edifice. The cyan lines represent the least squares fit
for the entire dataset. The black lines are the maximum uncertainties. The geometrical spreading and
the average Q obtained from the inversion are shown above each plot. Notice the progressive change
with increasing frequency. From De Siena et al. (2014a).

scatterers (De Siena et al., 2014b), using scattering tomography (also known as Nishigami (1991)
technique). The method can still be applied, but fixing the geometrical spreading will trade-off mea-
surements at low frequencies. Sketsiou et al. (2020) discusses the corresponding trade-offs in volcanic,
fault and regional settings. In Fig. 2.3 we show the same plot for the normalised S-wave energy
at Mount Vesuvius, where all sources and stations are inside the more heterogeneous domain of the

15
volcanic edifice and feeding system. Here, only data at 18 Hz can be used. Following De Siena et al.
(2014a), measurements above (below) the logarithmic linear trend shows high (low) attenuation. This
procedure is inverted by this version of MuRAT using a ray bending approach (Block, 1991; De Siena
et al., 2010), discussed computationally in Chapter ???.

Figure 2.3: Same as Fig. 2.2 for Mount Vesuvius, where earthquakes and stations are all comprised
under the volcanic edifice. Notice the negative Q at low frequencies. From De Siena et al. (2014a).

16
Chapter 3

Scattering attenuation: peak delays

Timeline in papers.

Paper Descriptor Where

(Saito et al., 2002) Markov approximation Link

(Takahashi et al., 2007) Peak-delay regionalisation Link

(Takahashi et al., 2009) Peak-delay tomography Link

(Calvet et al., 2013) Peak delay regionalisation + Qc: crust Link

(De Siena et al., 2016) Peak delay + Qc: volcano Link

(Sato, 2016) Peak delays - short and long λ Link

17
3.1 Parameters

Input Parameters for MuRAT

Symbol Descriptor Where

tmin minimum peak delay allowed Set in input file

tmax maximum peak delay allowed Set in input file

t0 (s) Origin time SAC Haeder

3.2 The peak delay method in pills

The method measures the delay of the maximum energy of the envelope of the direct way to model
scattering attenuation with regionalization.

Pros of measuring scattering attenuation using peak delay

1. Single-station measurements - you can get all the info from a single seismogram.

2. Proven sensitivity to structural information and geologic boundaries.

Cons of measuring scattering attenuation using peak delay

1. Assumptions underlying the Markov approximation could be unfulfilled. Tackled by MuRAT

estimating parameters at different frequecies.

2. No inversion, only regionalization. Still not tackled by MuRAT.

Figure 3.1 is a visual representation of the method.

18
Figure 3.1: The PD method measures the delay of the maximum of the envelope from either the
arrival time or the origin time of the event.

3.3 The theory bit

While travelling, direct wave packets broaden because of scattering and diffraction. The seminal
work for the application of peak delay (envelope broadening) imaging are Aki and Chouet (1975),
who demonstrate that coda of high-frequency seismograms originate from random inhomogeneity in
elastic properties, and (Sato, 1989), who first proposed peak-delays as a tool to characterise random
heterogeneity. Fehler et al. (2000) proved that the Markov approximation for the parabolic wave
equation can synthesise seismic waveforms in two dimensions, including early and intermediate coda.
The seminal work for the technique is Saito et al. (2002), who developed the equations for forward
modelling envelopes for spherically outgoing media characterised by von Kármán power spectral density
functions (PSDF).

Saito et al. (2002) provides both an historical view and a discussion of the assumptions underlying the
method:

• The primary assumption of the method is that the wavelength (λ = 2π/k) is smaller than the

19
 correlation distance (a), so that if this is multiplied by the wavenumber: ak >> 1. Obviously,
the estimation of a is challenging but it appears likely that the assumption will be broken at low
frequencies, where diffraction takes over.

• Propagation is entirely constrained within a single parallelepiped of average velocity V0 . The

velocity field depends on space (r) as:

V (r) = V 0(1 + ξ(r)) (3.1)

where ξ(r) << 1 are random velocity fluctuations. For a wider description of how to derive auto-
correlation and power spectral density functions describing fluctuation distributions refer to Sato
et al. (2012). These quantities are controlled by  (the root-mean-square of velocity fluctuations),
a (the correlation length) and κ (tuning the richness of short-wavelength components of the
random medium).

• Frequencies are above 1 Hz, as below that the wavelength grows to a level where most assump-
tions are invalid in lithospheric settings.

The original definition models envelope duration (tq ) as the quantity related to the characteristic time
of envelope broadening (tM ) (Saito et al., 2002). It is instructive to look at the definition of tM given
by Sato (2016):

2
tM (κ, ζ, kc , r0 ) = CL (κ, ζ, kc )r02 ; (3.2)
2V0 a
where ζ, which is a number varying between 0.25 and 1.75. Sato (2016) defines it as a tuning parameter
and separates random media between long- and short-scale components, so that the correlation distance
for small-scale component is : a−1
S = ζkc .

This quantity fully defines scattering in the case of long wavelength heterogeneities, depending on
average velocity (V0 ), hypocentral distance (r0 , which in a homogeneous medium is the product of V0
and travel time of the phase t), correlation distance (a), and mean squared velocity fluctuations (2 ),
ωc
the wavenumber, depending on the angular frequency (ωc ) as kc = V0 and where:

 1/2 1
 π Γ(κ+ 2 ) (1 − (ζak )1−2κ ) κ 6= 1
(2κ−1)Γ(κ) c 2
CL (κ, ζ, kc ) = (3.3)
2ln(ζak ) 1
c κ= 2

20
The dependency on V0 and a seem straightforward, at least until we do not make r0 and CL (κ, ζ, kc )
explicit. Let’s see what happens for κ = 21 :

1 2 2 V0 t2 ln(ζakc )
tM (κ = , ζ, kc , r0 ) = V02 t2 2ln(ζakc ) = ; (3.4)
2 2V0 a a

in which case:

• The characteristic time increases quadratically with travel time, hence the need for correcting
this dependency to focus on scattering.

• A sure way to increase scattering is to increase 2 -beware that these variations are generally
limited.

• Characteristic times increase with frequency and ζ although only logarithmically.

• For actual ranges of correlation distances, tM can change a lot.

3.3.1 Applicability

The important assumption for the applicability of peak-delay analysis is that scattering is produced
by long-scale components. To check the validity, one must always test the following conditions:
ζakc >> 1. In general, this means frequency above 1 Hz. In this case, we can rewrite:

Figure 3.2: Broadening for different κ values. For parameters used, see Sato (2016).

21
  1 1

 2
r 2 π 2 Γ(κ+ 2 ) (ζak )1−2κ κ< 1
 2V a (1−2κ)Γ(κ) c 2
 0

 2 1
2
tM (κ, ζ, kc , r) = 2V a r ln(ζakc ) κ= (3.5)
 0 2
 1 1
 2 r2 π 2 Γ(κ+ 2 )

1
κ>

2V0 a (2κ−1)Γ(κ) 2

This is probably the most important relationship as it shows the lack of changes with frequency for
1
κ> 2 and the slow dependency for κ = 12 . Characteristic times, envelopes and scattering are strongly
frequency-dependent when κ < 12 . This gives us an immediate test on our data. Do waveform broaden
dramatically at high frequencies (Fig. 3.2)? This is already an indication of the richness in long wave
numbers (short-scale components). The first check the user should do is then visual on the data.

3.4 MuRAT workflow

MuRAT works in analogy to Takahashi et al. (2007), Tripathi et al. (2010) and Calvet et al. (2013)
when mapping peak delays. These authors show that above 1 Hz one must check the relationship
between hypocentral distance and measured peak delay. In Figures 3.3-3.4, we show two examples
for different hypocentral distances. It is clear that there is a threshold over which you can apply the
approximation. Hence Murat shows you the same plot for your dataset - we will discuss in Chapter ??
the examples at Mount St. Helens, Vrancea and Toba. The peak-delay method requires the same ray
tracing approach used for the CN method, discussed computationally in Chapter ??.

Following Takahashi et al. (2007), measurements above (below) the logarithmic linear trend shows
high (low) scattering. This procedure is regionalised by this version of MuRAT. It means that positive
and negative variations are allocated via their average over blocks after raytracing. The variation is
weighted by the length of each segment. Takahashi et al. (2009) show that there are better procedure
to invert for the characteristics of power spectral density function using peak delay: we leave it for
future versions of the code.

22
Figure 3.3: Logarithmic plot of peak delay times against hypocentral distances. Black dots represent
the data used in this study. White lines are regression lines. From Takahashi et al. (2007).

Figure 3.4: Logarithmic plot of peak delay times (in seconds) as a function of the hypocentral distance
(in kilometers) for crustal S waves in four frequency bands. Gray dots are the data, and black lines
are the regression lines: log10(Tpd) = Ar(f) + Br(f)log10R. From Calvet et al. (2013).

23
24
Chapter 4

Absorption: coda attenuation

Timeline in papers.

Paper Descriptor Where

(Aki and Chouet, 1975) Origin of Coda Link

(Fehler et al., 1992) MLTWA Link

(Calvet et al., 2013) Peak delay regionalisation + Qc: crust Link

(De Siena et al., 2016) Peak delay + Qc: volcano Link

(Del Pezzo et al., 2016) 2D RTT sensitivity kernels: analytic Link

(De Siena et al., 2017a) 2D inversion with analytic kernels Link

(Del Pezzo et al., 2018) 3D RTT sensitivity kernels: diffusive Link

(Akande et al., 2019) 3D inversion with diffusive kernels Link

(Sketsiou et al., 2020) Absorption tomography Link

25
4.1 Parameters

Input Parameters for MuRAT

Symbol Descriptor Where

tk starting lapse time Set in input file

tw coda window length Set in input file

t0 (s) Origin time SAC Haeder

4.2 The Qc method in pills

The method measures the decay of the envelope from a given lapse time as a way to model absorption
with an inversion procedure.

Pros of measuring absorption using Qc

1. Single-station measurements - you can get all the info from a single seismogram.

2. Proven sensitivity to fluids.

3. High-resolution information on near-source and near-station structures.

4. Full inversion procedure.

Cons of measuring scattering attenuation using peak delay

1. The diffusive assumption is rarely fulfilled, with the onset of surface waves corrupting especially
low frequencies. Tackled by MuRAT estimating parameters at different frequecies.

26
 2. Assumption of single highly-diffusive layer rarely fulfilled. Still not tackled by MuRAT.

Figure 4.1 is a visual representation of the method.

Figure 4.1: The Qc method measures the decay of the envelope from a given lapse time from the
origin time of the event.

4.3 The theory bit

4.3.1 MLTWA

Coda waves can be used for more than just normalizing direct wave energies. Starting from the seminal
work of Aki and Chouet (1975), coda waves have been used to characterize the tectonic state of the
Earth. This has lead to the development of the Multiple Lapse Time Window Analysis (MLTWA)
−1
technique Fehler et al. (1992) to separate scattering attenuation Q−1
s from absorption Qi . It does
−1 Q−1
so by measuring extinction length (Le = Q−1
s + Qi ) and seismic albedo (B0 = Le ).
s

Two assumptions are necessary to apply the MLTWA technique: the direct S-wave dominates the early
portion of an S-wave seismogram; the S-coda comprises scattered S-waves. While in its first portion

27
the S-wave amplitude is controlled by the total attenuation of the medium, coda-waves are a product of
scattering (Sato et al., 2012). The filtered seismograms are divided into three time-windows, starting
from the S arrival, defined by a starting time tk and window length tw . The time integrals Ek (r) of
the seismic energy density E(r, t) in each window k is:

Z tk +∆t
Ek (r) = E(r, t)dt k = 1...3 (4.1)
tk

where the energy is still a function of the source-station distance r. The wave energy decay for each
cycle −∆E/E is a function of the total quality factor Q (related to the attenuation coefficient η) via
the following equation:

 
∆E 2π 1 1 υη
− = = 2π + = (4.2)
E Q Qi Qs f

where υ is the average wave speed and f is the frequency, respectively. The next step consists in the
correction of the energy of each window Ek (r) for both the geometrical spreading and the integral of
the tw -seconds-long normalisation window (the last window in time). This procedure removes source
and site effects (Mayeda et al., 1992):

!
Ek (r)
Ekobs (r) = log10 4πr R t +∆t2
(4.3)
coda
tcoda Ek (r, t)dt

The resulting logarithms are plotted versus distance in the four frequency bands considered and fitted
to the theoretical normalised energies. The theoretical curves are modelled using the L2-norm misfit
function for a number of waveforms from i to N :

N X
X 3
M (L−1
e , B0 ) = [Ekobs (ri ) − Ektheo (ri , L−1
e , B0 )]
2
(4.4)
i=1 k=1

Ektheo (ri , L−1

e , B0 ) is the theoretical normalised energy computed at distance ri for the k-th time-
window, fixing the L−1
e and B0 obtained from MLTWA. M is minimised with a grid search in the
B0 and L−1 −1
e parameter space, and the best-fit values L̂e and B̂0 are given by the minima of the
function. The isolines of the variable Mnorm = M (L−1 −1
e , B0 )/M (L̂e , B̂0 ) give the error estimates on
L̂−1
e and B̂0 . Being Mnorm an F-variable (Del Pezzo and Bianco, 2010), the model parameters can

28
be estimated using an F distribution (Pisconti et al., 2015), with a level of confidence higher than
F=0.68.

4.3.2 Sensitivity kernels

Due to the increased lateral sensitivity compared to ray-dependent techniques, coda waves are the-
oretically the ideal tool to image the Earth. In practice, apart for the cases of single scattering and
diffusion, there is generally no simple analytical solution, and envelopes have to be dealt with Radia-
tive Transfer Theory (Sato et al., 2012), possibly mixed with ray-tracing techniques that model direct
wave information (Sanborn et al., 2017). However, if the wave field is diffusive, one can obtain coda
sensitivity kernels with Paasschens equations (Paasschens, 1997) and the approximations of Pacheco
and Snieder (2005). Several authors have obtained 2D diffusive kernels based on Radiative Transfer
Theory (Obermann et al., 2013; Margerin, 2013; Mayor et al., 2014; Del Pezzo et al., 2016). These
kernels have maximum sensitivity at source and receiver (they show a pole at their exact location that
must be interpolated) and are all very similar (see e.g. Figs. 4.2) despite being obtained with different
approximations. They are primarily aimed at coda-wave interferometry but have found applications
in coda-attenuation imaging at regional and local scales (Mayor et al., 2016; De Siena et al., 2017a;
Napolitano et al., 2020; Sketsiou et al., 2020). (Calvet et al., 2013) first noticed the contribution of
surface waves to low-frequency observations, an effect modelled in 2D by Obermann et al. (2013) and
fit on data by Gabrielli et al. (2020). Del Pezzo et al. (2018) have extended the formulation to 3D
media, setting the spatial and frequency limits of the kernel applicability and discuss the assumptions
behind the applicability of the kernels.

There are two published approaches to map coda attenuation in the 3D space with kernels. The first is
regionalization (Del Pezzo et al., 2018) in analogy to peak-delay mapping. An excellent review paper
on this approach is Del Pezzo and Ibáñez (2020). MuRAT follows instead the pathway laid down by
De Siena et al. (2017a). It assumes that all coda energy is contained within the propagation grid at the
chosen lapse time. If this is true, then coda attenuation in space can be obtained from source-station
measurement with a standard inversion. Sketsiou et al. (2020) provides an extensive discussion of the
differences between regionalization and inversion. Akande et al. (2019) is the first 3D kernel-based
coda attenuation model obtained with 3D sensitivity kernels and an inversion approach.

As anticipated in Chapter 1, the combined effect of coherent and incoherent scattering as well as
the trade off induced by changing scattering regimes (from Rayleigh to Mie scattering, Cormier and
Sanborn (eg 2019)) can deeply affect the reliability of the kernels. In order for these to be applied,

29
 Figure 4.2: Diffusive sensitivity kernels at time 3.6 s obtained by Obermann et al. (2013)

there is a need for a volume with uniform (or gradually-changing) scattering properties, which is rarely
valid at all frequencies investigated. Surface Obermann et al. (2013) and resonant Margerin (2013);
De Siena et al. (2013) waves, sharp horizontal transitions in scattering (Wegler, 2005; Margerin, 2017;
Nardoni et al., 2021) and more generally not accounting properly for the trade-offs with geometrical
spreading (Morozov, 2011) all bring to inexact, if not incorrect, images of the Earth’s scattering and
absorption properties.

4.3.3 Applicability

Coda attenuation is measured by MuRAT in the diffusive assumption, hence one must check that
Q−1
c does not vary with hypocentral distance (Calvet et al., 2013). However, this only attests that
equipartition is in place, a necessary but insufficient condition for diffusion. The analysis of coda
attenuation with frequency is central to test the applicability of both MLTWA and sensitivity kernels
(Sketsiou et al., 2020). Also, as we are using the inversion procedure, it is central to test that most
of coda energy remains inside the inversion grid (Akande et al., 2019).

MuRAT offers tools to test these approximations and the effect of Qc on measurements of total

30
attenuation (chapter 2).

4.3.4 MuRAT workflow

MuRAT works in analogy to Mayor et al. (2016), De Siena et al. (2017a) and Akande et al. (2019)
when modelling coda attenuation. These authors show that above 1 Hz one must check the rela-
tionship between hypocentral distance and measured coda attenuation. In Figure 4.3, we show the
expected relation of coda attenuation versus hypocentral distance for the applicability of the theory
and interpretation of Qc as measurement of absorption. It is clear that the standard assumption of
twice the S-wave travel time (Aki and Chouet, 1975) does not work. Hence Murat shows you the
same plot for your dataset. However, the ideal threshold (100 s in this particular case) removes too
many data for reliable coverage. Figures 4.3c shows the acceptable compromise.

Figure 4.3: (a) Qc as a function of the epicentral distance in the 4–8 Hz frequency band. The duration
of the coda window is fixed (tw = 50 s). The solid and dashed lines show the mean dependence of Qc
with the epicentral distance and the associated uncertainties, respectively. The analysis is performed
on the onset of coda window tk equal to (a) 2tS (b) 100 s and (c) 70 s. From Mayor et al. (2016)

MuRAT also outputs the sensitivity kernels in two grids: (1) the fine propagation grid, extending
generally well beyond the area of inversion, and (2) the inversion grid. The first is used to compute
the full kernels and interpolate the poles at source and receiver locations (Akande et al., 2019). The
boundaries of the second are chosen by the user depending on source-station coverage. The code
shows what percentage of the energy is still contained inside the inversion grid.

31
Chapter 5

Getting started

5.1 Installation and set up in a nutshell

The current version of the code works following these steps:

1. Download or clone the package at the corresponding GitHub page;

2. Work in the downloaded folder after moving it to an appropriate location on your system;

3. Check that the IRTools have been downloaded in the corresponding folder in the working direc-
tory. Otherwise download them from the corresponding GitHub page;

4. Open one of the three input .mlx files, providing a step-by-step explanation of all inputs
(M urat_inputM SH.mlx, M urat_inputRomania.mlx, or M urat_inputT oba.mlx) and
create your own;

5. MuRAT works with SAC files that must be stored into a single folder and corrected for the
instrument function. The files must have populated headers, although the code can work using
just some header fields (see Chapter 6);

6. Run MuRAT3.mlx and select the name of the input file desired.

5.2 Matlab

MuRAT wants to be an open-access code, yet it also comes in Matlab format:

why?

• The first reason is historical. The code has been designed with core Matlab program, with
legacy bash and Fortran codes embedded in it. Passing from Matlab to Python, the obvious

32
 choice for open-access interfaces, was not straightforward. Passing from Matlab to Julia has
been way easier. Evolving through a decade, the Matlab version of MuRAT has reached the
format presented here. Yet, MuRAT3.0 will be the last release in Matlab.

• The second reason is simplicity of learning. MuRAT has been thought from start as a way to
introduce students to computational seismology. Matlab an ideal interface to teach computa-
tional codes to students in Earth Sciences. MuRAT has also been developed with the wider
geological community in mind, which until recently was often lacking training in computation.
The average geology student understands Matlab better than other codes.

• The third reason is that Python is becoming the standard in seismology, especially thanks to
tools like Obspy, yet this is not so obvious for the rest of Earth Sciences. For example, Matlab
is (has been) the favourite language interface for most of the Geodynamics community - see
for example Taras Gerya’s Introduction to Numerical Geodynamic Modelling. The geodynamics
community is now embracing Julia as its natural successor. MuRAT does the same.

5.2.1 Requirements and installations in pills:

• DOWNLOAD: MuRAT from Github (Fig. 5.1). Ideally, fork or star the code to get updates.
This manual explains how to work with GitHub Desktop.

• SYSTEM: The program works on Mac, Linux and Windows systems equipped with Matlab
R2019a or later.

• NECESSARY TOOLBOXES: Signal Processing, Curve Fitting, Image Processing and Mapping
Toolboxs. The Parallel Computing Toolbox is recommended for speed.

• EXAMPLE DATASETS: Three sample datasets (Mount St. Helens, Romania, and Toba) are
included in the package and allow the user to obtain sample models. The datasets work with
the input .mlx files that are provided in the working directory.

5.2.2 Changing folder

In this example, MuRAT is cloned from the corresponding page (Fig. 5.1) to a folder inside GitHub
Desktop. As a developer or user of MuRAT for multiple datasets, you might want to copy this folder on
your system at a different location (Fig. 5.2). In the example (Fig. 5.2,a), we show how you download

33
 Figure 5.1: The MuRAT page on GitHub.

automatically the main version of the code in a folder contained in the Github folder. Depending
on your system, the path to this folder could be different. It is recommended to copy and rename
the folder to another path (Fig. 5.2,b), and leave the original MuRAT folder where it is downloaded.
When finished with your work or if cooperating on that specific dataset with someone, you can then
substitute your edited folder in Github. This will allow you to push a new branch that might be used
as repository for codes and results. These repositories are now a requirement for most journals and
allow you to get the seminal version of your results in case of reviews.

5.2.3 README.md

The simplest way to start with MuRAT is reading its readme.txt. This files provides you with instruc-
tions in a nutshell on how to install the code, set up your data, workflow, and outputs. This is the
lowest level of documentation but covers the basics without going into jargon or trying solving all
problems at once.

34
Figure 5.2: How I work with MuRAT in Matlab. a) The folder where the branched code is downloaded
(left) vs the working directory (right). b) The Workspace in Matlab. Notice that, in theory, only the
Matlab input files are present in the folder and must be edited, if data are prepared correctly.

5.2.4 Wiki

Higher-level documentation is provided by the MuRAT Wiki (Fig. ??). As any wiki on GitHub, it has
a lateral panel that allows to reach pages that explain, for example, data preparation or input files.
The primary reason why the user should use MuRAT are the videos. Several pages present videos that
explain (Luca De Siena’s) way of working with the code. These videos might be the simplest, most
straightforward way to use the code.

5.3 Julia

MuRAT will be an open-access code developed in Julia from this version onward.

Julia is FAST. Julia is EASY. Julia is more similar to Matlab. Hence, the choice to switch from Matlab
to Julia. Is it better than Python? This is an absurd question for two languages developed in different
decades. If you need to call a Python script (and sometimes you do), you can do it from Julia. Please
have a read at this interesting comparison to know more about how the two codes compare.

35
Figure 5.3: MuRAT has a Wikipedia-like environment that provides specific documentation on tasks
and structures of the code.

36
Chapter 6

Data preparation

6.1 SAC files

The Seismic Analysis Code has been the standard in seismology for at least 20 years. In order to
process the SAC data in Matlab, we use the MatSAC tools created by Prof. Zhigang Peng from
Georgia Tech. MuRAT owes a lot to the format defined by Peng with his f get_sac function:

[t, data, SAChdr] = f get_sac(pathSACF ile) (6.1)

where, given the path of the SAC file, the user obtains two variables (time and data) and especially a
structure (SAChdr ) that stores the fields of the header.

This chapter deals with the original SAC files and how you must populate their header in order to
work with MuRAT. By populating the header, the user will need no additional file related to data
processing.

6.2 Parameters

The aim in MuRAT has always been to do attenuation imaging using SAC files that can be downloaded
from online servers. These data are provided via miniseed that must be unpacked. The resulting SAC
might have already populated headers: it is up to the user to populate the parameters shown in the
following table if they are blank (or -1234 as by SAC standard).

37
 Necessary Parameters in SAC header

SAC field Descriptor Name in Matsac

header

o (s) origin time SAChdr.times.o

a (s) P-wave pick SAChdr.times.a

t0 (s) S-wave pick SAChdr.times.t0

stla (deg) station latitude SAChdr.station.stla

stlo (deg) station longitude SAChdr.station.stlo

stel (m) station elevation SAChdr.station.stel

evla (deg) event latitude SAChdr.event.evla

evlo (deg) event longitude SAChdr.event.evlo

evdp (km) event depth SAChdr.event.evdp

38
6.3 sac_dataf older - checking data

The bold-faced parameters in the above table are mandatory. The user can check the parameters
in the header of each SAC file using the utilities M urat_test.m and M urat_testAll.m, in the
Utilities_Matlab folder. After running MuRAT the first time, this folder will be added to your path.
Otherwise just copy these scropts in your working directory folder.

6.3.1 M urat_test - testing single waveforms

Figure 6.1: Example run of Murat_test. The image is shown while the SACheader is created in the
workspace. All the necessary parameters are checked. Notice the missing S-wave picking - it is possible
that the variable storing this parameter is different from t0.

% function [image, SAChdr] = Murat_test(nameWaveform,...

% centralFrequencies,smoothingC,figOutput,verboseOutput)
% TEST seismogram envelopes for changes in broadening

39
% CREATES a figure with seismograms and envelopes for different frequencies
%
% Input Parameters:
% nameWaveform: name of the SAC file
% centralFrequencies: vector of frequencies (Hz),if [] no filter
% smoothingCoefficient: coefficient to smooth envelopes
% figOutput: decide if you want to show figures (set to 1)
% verboseOutput: decide if you want to show messages (set to 1)
%
% Output:
% image: image with envelope at specified frequency
% SAChdr: header of the SAC file
%

The function uses the path of a waveform, a vector of central frequencies, and a smoothing coefficient
as inputs. There are two additional flags, that allow to show the corresponding seismogram and
envelope (set f igOutput = 1) and the output messages (set verboseOutput = 1). It provides the
user with the header structure (SAChdr). The fields necessary to apply MuRAT can be checked in
the workspace and the command window, or displayed directly with the verboseOutput on. Figure 6.1
shows an example run of Murat_test for a file stored in the sac_MSH folder, filtered at 6 Hz with
smoothing of 8 samples per cycle.

6.3.2 M urat_testAll.m - creating a file to check all headers in dataset

MuRAT checks that all the parameters in the headers for a dataset are populated. It produces error
messages for the missing parameters, pointing at the corresponding waveform. Finally it outputs the
excel file DataHaeders.xsl in the results, providing an overview of all the required parameters. Fig.
6.2 shows this file after a run over the data folder sac_MSH. Missing parameters will be blank in the
file. The origin and S-wave peaking (searched for the SAC variables o and t0) are missing. These are
optional parameters if using P waves so the code will work anyway.

The function can be use before running MuRAT. It provides a DataHaeders.xsl file in the working
folder that the user can check to set the variables in the header.

40
Figure 6.2: Example file after running MuRAT for the MSH example. All the necessary parameters
are checked. Notice the missing S-wave picking - it is possible that the variable storing this parameter
is different from t0.

6.3.3 M urat_changeHdr.m - changing headers of single files

The previous analysis could discover some waveforms with missing or wrong header fields. The best
procedure is to change these headers direct in SAC; however, users might not be familiar with the
code, so M urat_changeHdr.m offers an alternative to change header fields:

f% function seism = Murat_changeHdr(newfolder)

% CHANGES header of a file
%
% Input Parameters:
% newFolder: folder where you save the changed file
%
% Output:
% seism: new seismogram
%

The only input is the path to a folder where the user will store the SAC with changed header. The
user selects the SAC file to be changed through a GUI. Then, in the command window, the user is
asked if the event, station or time fields are to be changed. Most often the problem is with the time.
The code assigns the origin to field 0 o0 , the P-time to 0 a0 , and the S-time to 0 t00 .

41
Chapter 7

Input files

MuRAT3 offers three input files as examples (Fig. 7.1). A new user needs to edit these files in order
to run the code. The details of the files are given into each mlx file. The Mount St. Helen input
file is shown in Appendix A. The new user can use this M urat_inputM SH.mlx file to learn what
each parameter does; however, users soon switch to a simpler .m file. The input files are divided into
sections, whose inputs are described in the following. For a more precise description, consult the .mlx
files or Appendix 1 in this Documentation.

7.1 The fields of the input

• GENERAL FIELDS:
M urat.input.dataDirectory: Define the name of folders where data are located
M urat.input.label: Defines the name of the folder where results will be stored
M urat.input.workers: Number of cores if parallelised, leave empty for sequential code

• WAVEFORM DATA
M urat.input.originT ime: Defines the variable that stores the origin time, or leave empty
M urat.input.P T ime: Defines the variable that stores the P-time
M urat.input.ST ime: Defines the variable that stores the S-time, or leave empty
M urat.input.P orS: Defines if the analysis is for P- or S-waves
M urat.input.centralF requency: Defines the set of central frequencies analyzed
M urat.input.components: Defines the number of components
M urat.input.declustering: Number used to divide the grid and decluster the dataset

• PEAK DELAY

42
Figure 7.1: Input files are provided as M urat_input..mlx files. There are three sample inputs (Mount
St. Helens - MSH, Romania and Toba caldera) in the standard release.

M urat.input.minimumP eakDelay: Minimum peak delay allowed

M urat.input.maximumP eakDelay: Maximum peak delay allowed

• DIRECT WAVE ATTENUATION

M urat.input.spectralDecay: Spectral decay of coda, for surface, diffusive or body waves
M urat.input.bodyW indow: Window to measure body waves and noise in seconds
M urat.input.startN oise: Seconds to start noise window from start of recording
M urat.input.tresholdN oise: Minimum signal-to-noise ratios allowed

• CODA ATTENUATION
M urat.input.lapseT imeM ethod: Defines the method used to measure coda waves

43
 M urat.input.startLapseT ime: Defines start of coda wave window
M urat.input.codaW indow: Defines the length of coda window
M urat.input.maxtravel: Set the max travel times after which you exclude traces
M urat.input.albedo: Measured Albedo from MLTWA
M urat.input.extinctionLength: Measured Extinction Length from MLTWA
M urat.input.kernelT reshold: Coefficient to make kernels more accurate or quicker to
compute
M urat.input.QcM easurement: Decide if using a ’Linearized’ or ’Non-Linear’ approach
M urat.input.f itT resholdLinear: Minimum fit accepted in the Linearized case

• GEOMETRY AND VELOCITY

M urat.input.origin: Defines the origin of the spatial grid
M urat.input.end: Defines the end of the spatial grid
M urat.input.gridLat: Number of nodes across latitude
M urat.input.gridLon: Number of nodes across longitude
M urat.input.gridZ: Number of nodes across depth
M urat.input.sections: Location of sections where you will get the maps
M urat.input.availableV elocity: Decides if a 1D (0) or 3D (1) model is available
M urat.input.namev: Defines the name of the velocity model
M urat.input.averageV elocityP : Defines the average P-wave velocity
M urat.input.averageV elocityS: Defines the average S-wave velocity

• INVERSION AND PLOTTING

M urat.input.inversionM ethod: Inverts with zero-orde Tikhonov or Iteratively
M urat.input.lCurve: Decides if selecting the damping during computation
M urat.input.lCurveQc: Damping parameters for different frequencies, Qc inversion
M urat.input.lCurveQ: Damping parameters for different frequencies, Q inversion
M urat.input.sizeCheck: Decides if doubling or making node space 4 times in checkerboard
M urat.input.highCheck: Decides the maximum inverse attenuation in the checkerboard

44
 M urat.input.lowCheck: Decides the minimum inverse attenuation in the checkerboard
M urat.input.spikeLocationOrigin: Origin of the spike block
M urat.input.spikeLocationEnd: End of the spike block
M urat.input.spikeV alue: Value of attenuation in the block

Each input is necessary, but the user will change repeatedly only some of them. Here, I give details
on each group.

7.1.1 GENERAL FIELDS

These are set at the beginning and never changed, except M urat.input.workers. The first run
should always be sequential (set it to []). The sequential run must be used to debug the code. While
MuRAT has several embedded error messages, it is often necessary to stop the run especially during
data processing (see M urat_data). The debug function in Matlab does not allow pinpointing the
incorrect waveform with a parallelised processing. However, it is recommended to perform the final
runs in parallelised mode: the parallelised processing reduces computational time from ∼ 15 s to ∼ 4
s for the MSH dataset.

7.1.2 WAVEFORM DATA

You can check the SAC header requirements in chapter 6. Once set, the header variable must be the
same for all files (e.g., variable ’a’ for P-wave picking or ’o’ for the origin). The number of components
1
is defied at the start. The code allows a single analysis for multiple frequencies with bandwidth 3
of the central frequency. Finally, the declustering variable divides the grid into smaller cells to select
the best earthquake waveforms in the cell, avoiding different repeated measurements that increase
residuals. It is recommended to test the code first without declustering.

7.1.3 PEAK DELAY

See chapter 3 for additional details. Setting the minimum and, especially, the maximum allows avoiding
tracking phases unrelated to the direct wave arrival. Sometimes tracking these phases, typically surface
waves generated by a strong impedance contrast at the surface, is useful to map changes in topography
and geomorphology (De Siena et al., 2016; Gabrielli et al., 2020).

45
7.1.4 DIRECT WAVE ATTENUATION

See chapter 2 for additional details. The spectral day has minimal effect on variations in space, but
it can affect absolute values of total attenuation. M urat.input.bodyW indow must be small enough
to still assume a ray approximation but large enough to mitigate changes in radiation pattern, the
primary trade-off of the coda-normalisation method.

M urat.input.startN oise depends on how much noise you have at the start of your recordings. The
code will measure noise from this point on the window set for body wave and use it to compute
signal-to-noise ratios. All waveforms below M urat.input.tresholdN oise are discarded.

7.1.5 CODA ATTENUATION

See chapter 4 for additional details. There are three methods to measure coda: using a constant
start time (’Constant’), the peak amplitude (’Peak’) and a number multiplied by the travel time
(’Travel’). The standard practice is to use the last option (e.g., 2tS ), however, coherent waves are
often still in the corresponding window (see chapter 8) (Calvet et al., 2013). The ’Peak’ is used
in when employing active data in volcanics, as in that case the waveform becomes rapidly diffusive
(Wegler, 2003).

If the user has no idea of albedo or extinction length, he/she can use published values for the area
or the crust. M urat.input.kernelT reshold must be set to find a compromise between accuracy of
the kernels and computational costs. Setting it to 1 solves the forward kernels on the same grid used
in the inversion: this is typically too coarse, e.g., to interpolate the poles of the kernes at source and
station. The minimum value is set to two, with 4/8 providing the most accurate results.

While the standard practice is to linearise the inversion by taking the logarithm of the envelope, the
user can choose a non-linear approach (Napolitano et al., 2020). In the first case, the minimum
accuracy is set with M urat.input.f itT resholdLinear, representing the Paerson coefficient relative
to a line. The advice is to set it to a very low value (e.g., 0.1) as this measurement will be used
anyway to weight the inversion. In the non-linear case, the weight is given by the PDF of the misfit.

7.1.6 GEOMETRY AND VELOCITY

The extension and number of nodes of the inversion grid are often changed throughout the analysis
because of the need for testing and avoiding spurious earthquakes. MuRAT inverts on a regular grid de-

46
fined in latitude , longitude and depth. If using a 1D model (M urat.input.availableV elocity=0) the
user can input depths in km, otherwise using a 3D model (M urat.input.availableV elocity=1) depths
are in meters. The text file of the 3D velocity model must replicate the structure of modvM SH.txt,
for the 1D that of iasp91.txt. Both files are inside the velocity_models folder. The user can set
average velocities for P and S waves for standard calculations.

7.1.7 INVERSION

The regtools (M urat.input.inversionM ethod = ’Tikhonov’) and IR Tools ( = ’Iterative’) are the
two packages included in the code and necessary to invert for direct and coda attenuation. After
the first run of MuRAT, a new folder (Test) will provide several figures that show either the residual
vs norm function or a cost function that must be minimised (Aster et al., 2005). By selecting the
corresponding parameter for each frequency, the user dampens. the inversion.

The other inputs define the multiplication factor of the checkerboard test relative to node spacing (2
or 4) and the values of the alternating checkerboard anomalies. The inputs for the spike test define
its location and value.

47
Chapter 8

Tests

The code provides test figures for the three methods. These are all in the Tests folder, created by the
code in the Results folder. These tests are divided into three categories:

• Analyses of method data and inversion quality;

• Curves that allow defining the best damping parameters for the inversion;

• Plots of the behaviour of coda attenuation vs frequency and velocity models.

8.1 Analyses

The Analyses figures (files ∗.analysis_f _Hz) are the most important output in MuRAT. They allow
assessing the quality of the data and inversion (except for the peak delay, which is only regionalised).

8.1.1 Peak delays

As described in Chapter 3, the hypotheses behind peak-delay imaging is that peaks increase vs travel
times. Figure 8.2 shows the P D_analysis_f requency figures for the three sample datasets.

These figures allow you to compare the differences (and problems) coming from peak-delay analysis in
different frequency bands. The Mount St Helens example (Figure 8.1) shows how the 3 Hz frequency
is affected by surface waves acting at short travel-times. This effect is due to unconsolidated materials
from the 1980 debris avalanche (Gabrielli et al., 2020). Analyses are only possible above 6 Hz.

The user needs to check the fit of the peak-delay behaviour to a line in order to apply the
regionalization procedure, as explained in Chapter 3.

48
Figure 8.1: Comparison between peak delays for the four frequencies and the Mount St Helens dataset.

8.1.2 Coda attenuation

As explained in Chapter 4, coda attenuation must be constant with increasing ray length
(Calvet et al., 2013) to interpret values of Qc as absorption. MuRAT outputs this simple variation
for a first-order check on the data. A constant Qc does not mean that the field is diffusive, but
indicates that equipartition is likely taking place. For better calculations, one can use the results of
an MLTWA analysis (Fehler et al., 1992; Del Pezzo et al., 2001; Del Pezzo, 2008). Codes like Qopen
(Eulenfeld and Wegler, 2016) precisely measure mean free path and transport man free path, providing
better constraints on the onset of diffusion. However, their results are generally equivalent to those of
MLTWA (van Laaten et al., 2021).

We perform the inversion using the external packages regtools (Hansen, 2007) and their evolution for

49
iterative regularisation, the IRTools (Gazzola et al., 2019). The Picard condition gives a first glance
at how many coda-attenuation parameters will be solved by the inversion in space (see Aster et al.
(2005) for a review).

8.1.3 Direct-wave attenuation

The coda-normalisation method (chapter 2) requires the use of direct waves that decay linearly
with travel time (hypocentral distance). The coda-normalisation analysis (Fig. 8.4) shows how
its normalised logarithm decays for increasing travel time (upper panel). The data (black circles)
are compared to the analytical forward with average attenuation measured on the data (red stars).
The middle panel shows the behaviour of direct energy Without normalisation and correction for the
velocity model. The inversion is tested with the usual Picard condition, showing that, for the chosen
parametrisation, more than half of the parameters is solved.

8.2 Damping

The inversion for Qc and Q requires to set damping and/or smoothing parameters for each frequency.
The L_curve∗ figures show the relation between residual and norm of the solution, if using the
regtools (Hansen, 2007). Nevertheless, it is more sensible to employ the IRTools (Gazzola et al.,
2019), as they offer much better stability for large inversion problems and allow visualising a cost
function considering both damping and smoothing (Rawlinson and Spakman, 2016). In Fig. 8.5,
these two plots and corresponding cost functions are shown for the MSH dataset at 6 Hz. The Qc
inversion is way more unstable, as hinted by the Picard condition (Fig. 8.3), as shown by the difference
between minimisation parameters (two orders of magnitude). The red rectangle shows that damping
and smoothing are basically the same for the Q inversion, which solves the majority of the model
parameters (Fig. 8.4, Picard condition). However, in both cases a minimum is reached.

50
Figure 8.2: Comparison between peak delays for the three sample datasets.

51
Figure 8.3: Coda analysis plot for the MSH dataset at 6 Hz. The upper panel shows that Qc is
constant with the ray length computed through ray tracing. The lower panel depicts the result of the
Picard condition (Aster et al., 2005): only half of the parameter models are solved under the weakest
hypothesis (blue and red lines cross around the 150th parameter). Only one-tenth of the parameter
models are solved if we also consider the uncertainty on data. σ are the singular values.

52
Figure 8.4: Analysis of the data and inversion results for the coda-normalisation inversion at MSH for
the 6 Hz frequency band. The upper panel shows the decay of the logarithmic energy ratio with travel
time for the data (black circle) and the theoretical behaviour with average attenuation. The middle
panel shows the dependence of logarithmic energy on hypocentral distance. The Picard condition
(bottom panel) gives a first glance at how many coda-attenuation parameters will be solved by the
inversion in space (see Aster et al. (2005) for a review).

53
Figure 8.5: Output of the IRTools showing the relation between data fit and either the solution nor
or its smoothness (upper panels, equivalent to the L_curves) for the MSH dataset at 6 Hz. The
lowermost panels show the corresponding cost function to be minimised. In the case of Q, the cost
minimisation parameter (red rectangle) is the same for both conditions.

54
Chapter 9

Rays and Kernels

While scattering and total attenuation variations are based on a ray-bending technique (Block, 1991),
MuRAT includes 3D sensitivity kernels (Del Pezzo et al., 2018) for modelling seismic absorption in
space.

9.1 Ray tracing

MuRAT ports the original Fortran code for ray bending from Block (1991) into Matlab. The code is a
classical ray bending, starting from the locations of source and receiver and bending the ray in a way
that minimises travel time inside the input velocity model. The user can have a look at the function
M urat_tracing, which traces the ray in the velocity model. At the start of the function, there are
two optional parameters that set the maximum number of bending iterations and points allowed.

Figure 9.1: Rays of the coda-normalisation (a) and peak delay (b) methods at 3 Hz for the MSH
dataset.

55
Their values (maxit = 100 and maxpoints = 10000, respectively) have been set after several trials and
generally work just fine at the crustal scale. The rest of the procedure includes original function that
average velocities on the regular grid and calculate velocity gradients, informing the bending process
(Block, 1991). The functions have been tested against standard Matlab functions, and are generally
from three to ten times faster.

MuRAT pairs to the ray-bending approach some utilities optimised for speed that measure ray lengths
across the inversion grid. These lengths are used both in the standard coda-normalisation method
and as a weight in the peak-delay regionalisation (Chapters 2-3). The user can look at the func-
tion M urat_rays, which calls both procedures, to see that the procedure needs as input the ve-
locity model, a propagation grid, and source/station locations. Within the function, MuRAT calls
M urat_segments, calculating segment lengths that it includes directly in the coda-normalisation
matrix (see chapter 2) and peak-delay matrix. The corresponding rays are finally shown as result in
the folder RaysKernals. As the data selection is different for coda-normalisation and peak-delay, the
code outputs two different sets of Rays_* figures (Fig. 9.1). The ray tracing works for both the small
volcanic scale (like at MSH, Fig. 9.2a), mantle scale (Fig. 9.2b) and continental-oceanic settings (Fig.
9.2c).

9.2 Kernel implementation

MuRAT implements the computational kernels described by Del Pezzo et al. (2018) in the tomographic
procedure described by De Siena et al. (2017a). The kernels (Fig. 9.3) are calculated with a call to
the function M urat_kernels, using as input the albedo and extinction lengths input in the code.
Panel a) shows the full computational kernel computed up to a laspe time equal to the starting lapse
time plus half of the coda window.

56
Figure 9.2: Rays of the coda-normalisation method at 3 Hz for the three sample datasets: a) MSH;
b) Romania; c) Toba.

57
Figure 9.3: Kernels for coda-attenuation imaging at 3 Hz for the MSH dataset. The full kernels are
shown as a section in the 3D space (a) and in the grid chosen for the inversion (b). Notice that
sensitivity is always highest at source and receiver. The colour scale is logarithmic. The lapse time is
22.5 s.

58
Appendices

59
Appendix A

MuRAT input file - Mount St Helens

60
Table of Contents
GENERAL FIELDS.................................................................................................................................................. 1
WAVEFORM DATA..................................................................................................................................................3
PEAK DELAY........................................................................................................................................................... 4
DIRECT WAVE ATTENUATION..............................................................................................................................4
CODA ATTENUATION.............................................................................................................................................5
Lapse times and sensitivity kernels..................................................................................................................... 5
Measurement of Qc............................................................................................................................................. 6
GEOMETRY AND VELOCITY................................................................................................................................. 6
INVERSION.............................................................................................................................................................. 7

INPUT MuRAT3D - MSH

This is an input file for the program Multi-Resolution Attenuation Tomography (MuRAT), version 3. It refers to
the following area:
MOUNT ST HELENS VOLCANO

The working directory is the folder where you downloaded MuRAT: move it anywhere in your system.

GENERAL FIELDS
The user needs a data directory containing .sac files (beware of the difference between .sac and .SAC). If
you want to use two or three-component recordings, you have to store them into a single folder, but the three
components must be in the exact order E,N,Z. For an example see Toba. For this example we will use data in
the sac_MSH folder. Inside this folder, only vertical (Z) seismograms are stored:

Murat.input.dataDirectory = 'sac_MSH';

1
Best practice is to create a new folder to store your results. The code will create this folder inside the working
directory. Specify the name of the folder that will store text files and figures, and will appear in your working
directory:

Murat.input.label = 'MSH';

2
In MuRAT3D you can choose between a sequential or parallelized forward loop. In the parallelized case, just
set a number of workers (cores). In this example, we are working with the parallelized code and a computer with
8 cores. Otherwise, set Murat.input.workers as empty ([]).

Murat.input.workers = [];

WAVEFORM DATA
Set all the variables required by data processing. This includes data choises, as setting the name of the
variables in SAC and all the attributes that are needed for the three kinds of imaging. The routines for loading
the files have been mostly created by Zhigang Peng and co-workers and downloaded from their Introduction to
SAC webpage.

MuRAT3D is designed to work with sac files only, so it is necessary to set the name of the variable containing
their peakings and zero time. Files have to have populated headers. Set the name of the variable containing the
origin time, P-wave peaking and S-wave pickings - in this example there is only the mandatory P-wave picking.
If you don't have the origin time and S-wave times at hand you can mark them as empty ([]):

Murat.input.originTime = [];
Murat.input.PTime = 'a';
Murat.input.STime = [];

SAChdr is a structure that contains all the fields you saved in the SAC haeder. In this example the P-wave
picking has been stored inside variable a. You always find these variables in the times subfield:

3
Then choose the coherent phase you are analyzing - P-(2) or S-(3). In our case it is P- as we have no S-wave
picking:

Murat.input.POrS = 2;

You need to set the central frequencies (Hz) according to your spectrograms. General practice is to vary it
across your spectra (see De Siena et al. 2016, EPSL) for absorption and scattering mapping or focus on a given
frequency (De Siena et al. 2014, JGR) for direct-wave attenuation imaging. Here, they cover the interval [1.5-24]
Hz:

Murat.input.centralFrequency = [3 6 12 18];

You can work with 1 vertical or horizontal (1), 2 horizontal (2) or three components(3). If using more than one
component, the order MUST BE: WE, SN, Vertical or SN, WE, Vertical. In this example we work with the vertical
component only:

Murat.input.components = 1;

Finally, you can opt to decluster your data events. The code will divide the inversion grid by the following factor
and select the best earthquake located in the block among all others. Set it to empty if you want to opt out ([]).

Murat.input.declustering = 5;

PEAK DELAY
Scattering (peak delay) measurements rely on the existence of coherent waves. Peak delay is a standard
measurements of forward scattering in regional scale-mapping since Takahashi et al. 2007, GJI. Here, we
approach the problem by adding a ray-tracing strategy to the system, assuming that the sensitivity follows the
seismic ray. Also, we bring it to 3D and apply it to P-waves (see De Siena et al. 2016, EPSL).

To do so we need to set a minimum peak delay (s) considering scattering in the area and frequency. If using P
wave this is important, as they can be sometimes more energetic than S-waves, biasing our measurements. We
also need to set the maximum peack-delay (s) to avoid picking surface waves.

Murat.input.minimumPeakDelay = 0.1;
Murat.input.maximumPeakDelay = 7;

DIRECT WAVE ATTENUATION

Total attenuation (inverse Q) measurements also relies on the existence of coherent waves. Total attenuation
with the coda-normalization method is today astandard in volcano tomography (Del Pezzo et al. 2006, PEPI, De
Siena et al. 2010, JGR; De Siena et al. 2014, JGR, Prudencio et al. 2015, Prudencio & Manga, 2020.

The method relies on the measurements of both direct and coda wave energy. Therefore we need to set the
spectral energy decay of the coda wavefield. The spectral decay has to be set for prdominance of 2D surface
(0.5), diffusive (1.5) or body waves (2). Here it is it is the first case:

Murat.input.spectralDecay = 0.5;

4
Then we set the length of the window used to measure direct-wave energy, trying to smooth radiation pattern
effects. The code uses the lame length to measure noise energy. Discussions on the topic can be found in De
Siena et al. 2009, PEPI and De Siena et al. 2010, JGR, but the choise is as usual dependent on data:

Murat.input.bodyWindow = 1;

We also need the start of the window used to measure noise a few seconds after the start of the recording (in
seconds).

Murat.input.startNoise = 5;

The coda-to-noise energy ratio is used by the weighted inversion. Here you set the minimum accepted energy
ratio.

Murat.input.tresholdNoise = 3;

CODA ATTENUATION
Coda attenuation (inverse Qc) is measured from the decay of the coda with lapse time from the origin time of
the earthquake, and requires energetic scattering. Qc is a well know parameter for assesing tectonic structures
(Sato et al. 2012, Springer). In recent years it has been used as an imaging attribute at the regional (Calvet
et al. 2013, Tectonophysics; Borleanu et al. 2017, Tectonophysics), fault (Napolitano et al. 2020; Sketsiou et
al. 2020, PEPI) and, especially, volcanic scales (Prudencio et al. 2013a, GJI; Prudencio et al. 2013b GJI; De
Siena et al. 2016, EPSL; De Siena et al. 2017, GRL; Gabrielli et al. 2020, GJI).

Lapse times and sensitivity kernels

In MuRAT3D, we use the full 3D computational kernels devised by Del Pezzo et al. 2018, Geosciences in the
inversion approach proposed first by De Siena et al. 2017, GRL.

Here you choose the start of the window used to measure coda wave energy and model kernels. The starting
time of the coda window can be set directly in seconds ('Constant'), from the envelope peak ('Peak'), or
depending on the P- or S-wave travel time, for example twice the S-wave travel time ('Travel'). If the chosen
method is Constant, set start and length of the window.

Murat.input.lapseTimeMethod = 'Constant';

If the chosen method is Constant, set the start of the window in seconds after the origin time. It it is Peak, set
it to []. If it is Travel select Murat.input.startLapseTime as the moltiplicative factor of the pahase you are using
(e.g., 2 or 3). Avoid this method if you do not know the S-wave time.

Murat.input.startLapseTime = 15;

Finally set the length of the coda window in seconds. The true lapse time at which we calculate the kernels is
half of the window. The window is also used (after normalizing for its length) in the coda normalization method.

Murat.input.codaWindow = 15;

Set maximum travel time if you want to exclude traces beyond a certain value, else put [Inf].

5
 Murat.input.maxtravel = 13;

The MLTWA is the standard method to find the average parameters necessary to calculate the kernels. It
provides albedo and extinction length:

Murat.input.albedo = 0.5;
Murat.input.extinctionLength = 0.02;

As the kernels are computational intensive and require a full matrix of nodes to avoid singularities, we also
use a computational factor to reduce the computational time. This number divides the input grid, meaning that
higher numbers give more precise results at the expene of computational time. Minimum is 1. A figure will
output the kernel in this grid.

Murat.input.kernelTreshold = 2;

Measurement of Qc
MuRAT3D implements either a linearised approach or a grid search approach to measure Qc. The linearised
approach is the standard proposed first by Aki (e.g., Havskov et al. 2016, BSSA) to best fit Qc after taking the
logarithm of the energy. The uncertainties are derived from the simple minimum R-squared (fitTresholdLinear)
and needs to be defined by a number between 0 and 1. It is advisable to set a minimum of 0.1.

The non linear approach models energy data measured on one-second windows across the envelope and
minimizes the difference between data and model with a 1D grid search algorithm (Napolitano et al. 2020).
Uncertainties are given by the experimental probability density function of the misfit. In both cases, uncertainties
play as a weight in the final inversion. In the second case, leave the fitTresholdLinear = [].

The user needs to choose between the two options 'Linearized' and 'NonLinear':

Murat.input.QcMeasurement = 'Linearized';
Murat.input.fitTresholdLinear = 0.1;

GEOMETRY AND VELOCITY

This section sets the details of the inversion grid and availability of velocity model. In MuRAT3D the coordinates
of the model are in lat/lon, then they get converted in km. The vertical is in altitude above sea level. The velocity
model can be 1D or 3D - if 3D all points must be given in lat/long formats. You start by setting the origin and end
points of your inversion grid.

Murat.input.origin = [45.9805 -122.5030 3350];

Murat.input.end = [46.3900 -122.0000 -22650];

Then you need to set the number of nodes in the three directions. This is obviously dependent on the scale of
your area. You will be playing a lot with this to test your effective resolution capabiities.

Murat.input.gridLat = 10;
Murat.input.gridLong = 12;
Murat.input.gridZ = 6;

6
MuRAT3D now saves everything in Paraview format (.VTK) but some checks in Matlab can be useful. All these
sections are saved in the Label folder. You will see all of your figures on three sections cutting the models WE
(degrees), SN (degrees), and horizontally (meters or km) at:

Murat.input.sections = [46.12 -122.20 -1000];

With this version of the code you are always using an underlying velocity model: the 3D is either unavailable(0)
or a vailable (1) velocity model. For the 1D case MuRAT provides you iasp91.txt, the standard IASPEI velocity
model and expands it to a false 3D. However, a standard crustal model is generally available everywhere on
the Earth, so use that - but change it to the same format as the file provided, first column is depth, second is
distance from the centre of the Earth, then third and fourth are P- and S-wave velocity. Store the file in the
folder velocity_models. See the Romania example file. At Mount St. Helens, we use the 3D local earthquake
tomography model of Waite and Moran, 2009,JVGR, (modvMSH.txt), stored in the folder velocity_model:

whose coordinates are Lat/Long/Altitude/Velocity in meters. So we set:

Murat.input.availableVelocity = 1;
Murat.input.namev = 'modvMSH.txt';

Even if we set the velocity model we still need the average crustal velocities if you have no info of origin time. It
is highly recommended you have the origin in the haeder, at variables 'o'!

Murat.input.averageVelocityP = 6;
Murat.input.averageVelocityS = 3;

INVERSION
The code implements either a standard Tikhonov inversion based on singular value decomposition ('Tikhonov')
and an iterative conjugate graduate least square inversion minimizing both model norm and Laplacian
('Iterative'). For the first, we rely on the regtools Matlab suite from Per Christian Hansen. The second uses the
IR TOOLS for iterative regularization. Copies of both packages are linked to MURAT3:

Murat.input.inversionMethod = 'Iterative';

As you need to select the damping you can choose to output the L curves between residual and norm length
(Aster et al. 2013 - case 'Tikhonov') or the cost functions (case 'Iterative'). Set either 1 or 0. In the latter case,
set the damping for both Qc and Q inversions. You will likely test many different damping parameters and likely
want to avoid seeing the same figure again, so you can just declare the damping for Qc and Q after looking at
the figures. Remember doing it for each frequency.

Murat.input.lCurve = 0;
Murat.input.lCurveQc = [0.1 0.1 0.05 0.02];

7
 Murat.input.lCurveQ = [1 1 0.5 0.2];

A great reference for the best sort of testing is Rawlinson & Spakman, 2016. If you want to test you results you
need to create a checkerboard. The size of the checks can be twice (2) or four times (4) node spacing:

Murat.input.sizeCheck = 2;

with each cell having alternating values of attenuation:

Murat.input.highCheck = 0.02;
Murat.input.lowCheck = 0.001;

In MuRAT3D you can also set the origin and end locations of a spike as well as its attenuation value:

Murat.input.spikeLocationOrigin = [46.05 -122.40 0];

Murat.input.spikeLocationEnd = [46.12 -122.30 -6000];
Murat.input.spikeValue = 0.02;

8
Bibliography

W. G. Akande, L. De Siena, and Q. Gan. Three-dimensional kernel-based coda attenuation imaging

of caldera structures controlling the 1982-84 campi flegrei unrest. Journal of Volcanology and
Geothermal Research, 381:273–283, 2019.

K. Aki. Attenuation of shear-waves in the lithosphere for frequencies from 0.05 to 25 Hz. Physics of
the Earth and Planetary Interiors, 21:50–60, 1980.

K. Aki and B. Chouet. Origin of coda waves: Source, attenuation, and scattering effects. Journal of
Geophysical Research, 80:3322–3342, 1975.

R. Aster, B. Borchers, and C. Thurber. Parameter Estimation and Inverse Problem, volume 90 of
International Geophysics Series, Advances in Geophysics. Elsevier, Amsterdam, 2005.

L. V. Block. Joint hypocenter–velocity inversion of local earthquake arrival time data in two geothermal
regions. Ph.d. dissertation, Massachusetts Institute of Technology, Cambridge, 1991.

M. Calvet and L. Margerin. Lapse time dependence of coda Q: anisotropic multiple-scattering models
and application to Pyrenees. Bulletin of the Seismological Society of America, 103(3):1993–2010,
2013. doi: 10.1785/0120120239.

M. Calvet, M. Sylvander, L. Margerin, and A. Villaseñor. Spatial variations of seismic attenuation and
heterogeneity in the Pyrenees: Coda Q and peak delay time analysis. Tectonophysics, 608:428–439,
2013.

V. F. Cormier and C. J. Sanborn. Trade-offs in parameters describing crustal heterogeneity and intrinsic
attenuation from radiative transport modeling of high-frequency regional seismograms. Bulletin of
the Seismological Society of America, 109(1):312–321, 2019.

L. De Siena, E. Del Pezzo, F. Bianco, and A. Tramelli. Multiple resolution seismic attenuation imaging
at Mt. Vesuvius. Physics of the Earth and Planetary Interiors, 173:17–32, 2009.

L. De Siena, E. Del Pezzo, and F. Bianco. Campi Flegrei seismic attenuation image: evidences of
gas resevoirs, hydrotermal basins and feeding systems. Journal of Geophysical Research, 115(B0):
9312–9329, 2010.

69
L. De Siena, E. Del Pezzo, C. Thomas, A. Curtis, and L. Margerin. Seismic energy envelopes in volcanic
media: in need of boundary conditions. Geophysical Journal International, 192(1):326–345, 2013.

L. De Siena, C. Thomas, and R. Aster. Multi-scale reasonable attenuation tomography analysis (Mu-
RAT): An imaging algorithm designed for volcanic regions. Journal of Volcanology and Geothermal
Research, 277:22–35, 2014a.

L. De Siena, C. Thomas, G. P. Waite, S. C. Moran, and S. Klemme. Attenuation and scattering

tomography of the deep plumbing system of Mount St. Helens. Journal of Geophysical Research:
Solid Earth, 119(11):8223–8238, 2014b.

L. De Siena, M. Calvet, K. J. Watson, A. Jonkers, and C. Thomas. Seismic scattering and absorption
mapping of debris flows, feeding paths, and tectonic units at Mount St. Helens volcano. Earth and
Planetary Science Letters, 442:21—31, 2016.

L. De Siena, A. Amoruso, E. D. Pezzo, Z. Wakeford, M. Castellano, and L. Crescentini. Space-weighted

seismic attenuation mapping of the aseismic source of Campi Flegrei 1983?1984 unrest. Geophysical
Research Letters, pages n/a–n/a, 2017a. ISSN 1944-8007. doi: 10.1002/2017GL072507. URL
http://dx.doi.org/10.1002/2017GL072507. 2017GL072507.

L. De Siena, G. Chiodini, G. Vilardo, E. Del Pezzo, M. Castellano, S. Colombelli, N. Tisato, and

G. Ventura. Source and dynamics of a volcanic caldera unrest: Campi flegrei, 1983–84. Scientific
reports, 7(1):1–13, 2017b.

E. Del Pezzo. Seismic wave scattering in volcanoes. In Earth Heterogeneity and Scattering Effects
of Seismic Waves, volume 50 of Advances in Geophysics, chapter 13, pages 353–369. Elsevier,
Amsterdam, 2008.

E. Del Pezzo and F. Bianco. Two-layer earth model corrections to the MLTWA estimates of intrinsic-
and scattering-attenuation obtained in a uniform half-space. Geophysical Journal International, 182
(2):949–955, 2010.

E. Del Pezzo and J. M. Ibáñez. Seismic coda-waves imaging based on sensitivity kernels calculated
using an heuristic approach. Geosciences, 10(8):304, 2020.

E. Del Pezzo, F. Bianco, and G. Saccorotti. Separation of intrinsic and scattering q for volcanic tremor:
an application to etna and masaya volcanoes. Geophysical Research Letters., 28(16):3083–3086,
2001.

70
E. Del Pezzo, F. Bianco, L. De Siena, and A. Zollo. Small scale shallow attenuation structure at Mt.
Vesuvius. Physics of the Earth and Planetary Interiors, 157:257–268, 2006.

E. Del Pezzo, J. Ibañez, J. Prudencio, F. Bianco, and L. De Siena. Absorption and scattering 2-D
volcano images from numerically calculated space-weighting functions. Geophysical Journal Inter-
national, 206(2):742–756, 2016.

E. Del Pezzo, A. De La Torre, F. Bianco, J. Ibanez, S. Gabrielli, and L. De Siena. Numerically

calculated 3d space-weighting functions to image crustal volcanic structures using diffuse coda
waves. Geosciences, 8(5):175, 2018.

M. D. P. Di Martino, L. De Siena, D. Healy, and S. Vialle. Petro-mineralogical controls on coda

attenuation in volcanic rock samples. Geophysical Journal International, 2021.

T. Eulenfeld and U. Wegler. Measurement of intrinsic and scattering attenuation of shear waves in two
sedimentary basins and comparison to crystalline sites in germany. Geophysical Journal International,
205(2):744–757, 2016.

M. C. Fehler, P. Roberts, and T. Fairbanks. A temporal change in coda wave attenuation observed
during an eruption of Mount St. Helens. Journal of Geophysical Research, 93:4367–4373, 1988.

M. C. Fehler, M. Hoshiba, H. Sato, and H. Obara. Separation of scattering and intrinsic attenuation for
the Kanto-Tokai region, Japan, using mesurements of S-wave energy versus hypocentral distance.
Geophysical Journal International, 108:787–800, 1992.

M. C. Fehler, H. Sato, and L. J. Huang. Envelope broadening of outgoing waves in 2D random

media: a comparison between the Markov approximation and numerical simulations. Bulletin of the
Seismological Society of America, 90:914–928, 2000.

S. Gabrielli, L. De Siena, F. Napolitano, and E. Del Pezzo. Understanding seismic path biases and
magmatic activity at mount st helens volcano before its 2004 eruption. Geophysical Journal Inter-
national, 222(1):169–188, 2020.

S. Gazzola, P. C. Hansen, and J. G. Nagy. Ir tools: a matlab package of iterative regularization

methods and large-scale test problems. Numerical Algorithms, 81(3):773–811, 2019.

P. C. Hansen. Regularization tools version 4.0 for matlab 7.3. Numerical algorithms, 46(2):189–194,
2007.

71
P. Ho-Liu, H. Kanamori, and R. W. Clayton. Applications of attenuation tomography to imperial
valley and coso-indian wells region, southern california. Journal of Geophysical Research, 93(B9):
10501–10520, 1988.

L. Margerin. Diffusion approximation with polarization and resonance effects for the modelling of
seismic waves in strongly scattering small-scale media. Geophysical Journal International, 192(1):
326–345, 2013.

L. Margerin. Breakdown of equipartition in diffuse fields caused by energy leakage. The European
Physical Journal Special Topics, 226(7):1353–1370, 2017.

K. Mayeda, S. Koyanagi, M. Hoshiba, K. Aki, and Y. Zeng. A comparative study of scattering,

intrinsic and coda q for hawaii, long valley, and central california between 1.5 and 15.0 hz. Journal
of Geophysical Research, 97:6643–6659, 1992.

J. Mayor, L. Margerin, and M. Calvet. Sensitivity of coda waves to spatial variations of absorption
and scattering: radiative transfer theory and 2-d examples. Geophysical Journal International, page
ggu046, 2014.

J. Mayor, M. Calvet, L. Margerin, O. Vanderhaeghe, and P. Traversa. Crustal structure of the Alps as
seen by attenuation tomography. Earth and Planetary Science Letters, 439:71–80, 2016.

I. B. Morozov. Geometrical attenuation, frequency dependence of q, and the absorption band problem.
Geophysical Journal International, 175(1):239–252, 2008.

I. B. Morozov. Mechanisms of geometrical seismic attenuation. Annals of Geophysics, 54(3), 2011.

F. Napolitano, L. De Siena, A. Gervasi, I. Guerra, R. Scarpa, and M. La Rocca. Scattering and absorp-
tion imaging of a highly fractured fluid-filled seismogenetic volume in a region of slow deformation.
Geoscience Frontiers, 11(3):989–998, 2020.

C. Nardoni, L. De Siena, F. Cammarano, F. Magrini, and E. Mattei. Modelling regional-scale attenua-

tion across italy and the tyrrhenian sea. Physics of the Earth and Planetary Interiors, page 106764,
2021.

K. Nishigami. A new invertion method of coda waveforms to determine spatial distribution of coda
scatterers in the crust and uppermost mantle. Geophysical Research Letters, 12(18):2225–2228,
1991.

72
A. Obermann, T. Planes, E. Larose, C. Sens-Schönfelder, and M. Campillo. Depth sensitivity of
seismic coda waves to velocity perturbations in an elastic heterogeneous medium. Geophysical
Journal International, 194(1):372–382, 2013.

J. C. L. Paasschens. Solution of the time-dependent Boltzmann equation. Physical Review, 56(1):

1135–1141, 1997.

C. Pacheco and R. Snieder. Time-lapse travel time change of multiply scattered acoustic waves. The
Journal of the Acoustical Society of America, 118(3):1300–1310, 2005.

A. Pisconti, E. Del Pezzo, F. Bianco, and S. de Lorenzo. Seismic q estimates in umbria marche
(central italy): hints for the retrieval of a new attenuation law for seismic risk. Geophysical Journal
International, 201(3):1370–1382, 2015.

J. Prudencio, J. M. Ibáñez, A. García-Yeguas, E. Del Pezzo, and A. M. Posadas. Spatial distribution

of intrinsic and scattering seismic attenuation in active volcanic islands–ii: Deception island images.
Geophysical Journal International, 195(3):1957–1969, 2013.

J. Prudencio, L. De Siena, J. Ibáñez, E. Del Pezzo, A. García-Yeguas, and A. Díaz-Moreno. The

3D Attenuation Structure of Deception Island (Antarctica). Surveys in Geophysics, 36(3):371–390,
2015.

N. Rawlinson and W. Spakman. On the use of sensitivity tests in seismic tomography. Geophysical
Journal International, 205(2):1221–1243, 2016.

T. Saito, H. Sato, and M. Ohtake. Envelope broadening of spherically outgoing waves in three-
dimensional random media having power law spectra. Journal of Geophysical Research, 107(B5):
2089–2103, 2002.

C. J. Sanborn, V. F. Cormier, and M. Fitzpatrick. Combined effects of deterministic and statistical

structure on high-frequency regional seismograms. Geophysical Journal International, 210(2):1143–
1159, 2017.

H. Sato. Broadening of seismogram envelopes in the randomly inhomogeneous lithosphere based on

the parabolic approximation: southeastern Honshu, Japan. Journal of Geophysical Research, 94:
17735–17747, 1989.

H. Sato. Envelope broadening and scattering attenuation of a scalar wavelet in random media having
power-law spectra. Geophysical Journal International, 204(1):386–398, 2016.

73
H. Sato, M. C. Fehler, and T. Maeda. Seismic Wave Propagation and Scattering in the heterogeneous
Earth: Second Edition. Springer, New York, USA, 2012.

P. Sketsiou, F. Napolitano, A. Zenonos, and L. De Siena. New insights into seismic absorption imaging.
Physics of the Earth and Planetary Interiors, 298:106337, 2020.

P. Sketsiou, L. De Siena, S. Gabrielli, and F. Napolitano. 3-d attenuation image of fluid storage and
tectonic interactions across the pollino fault network. Geophysical Journal International, 226(1):
536–547, 2021.

T. Takahashi, H. Sato, T. Nishimura, and K. Obara. Strong inhomogeneity beneath Quaternary

volcanoes revealed from the peak delay analysis of S-wave seismograms of microearthquakes in
northeastern Japan. Geophysical Journal International, 168(1):90–99, 2007.

T. Takahashi, H. Sato, T. Nishimura, and K. Obara. Tomographic inversion of the peak delay times
to reveal random velocity fluctuations in the lithosphere: method and application to northeastern
japan. Geophysical Journal International, 178(3):1437–1455, 2009.

S. Takemura, T. Furumura, and T. Saito. Distortion of the apparent s-wave radiation pattern in
the high-frequency wavefield: Tottori-ken seibu, japan, earthquake of 2000. Geophysical Journal
International, 178(2):950–961, 2009.

J. N. Tripathi, M. Sato, and M. Yamamoto. Envelope broadening characteristics of crustal earthquakes

in northeastern Honshu, Japan. Geophysical Journal International, 182(2):988–1000, 2010.

M. van Laaten, T. Eulenfeld, and U. Wegler. Comparison of multiple lapse time window analysis and
qopen to determine intrinsic and scattering attenuation. Geophysical Journal International, 228(2):
913–926, 2021.

U. Wegler. Analysis of Multiple Scattering at Vesuvius Volcano, Italy, using Data of the TomoVes
active seismic experiment. Journal of Volcanology and Geothermal Research, 128:45–63, 2003.

U. Wegler. Diffusion of seismic waves in layered media: boundary conditions and analytic solutions.
Geophysical Journal International, 163:1123–1135, 2005.

U. Wegler and B. G. Luehr. Scattering behaviour at Merapi Volcano (Java) revealed from an active
seismic experiment. Geophysical Journal International, 145:579–592, 2001.

74
K. Yoshimoto, H. Sato, and M. Ohtake. Frequency-Dependent Attenuation of P and S Waves in the
Kanto Area, Japan, based on the Coda-Normalization Method. Geophysical Journal International,
114:165–174, 1993.

75