See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/232404428

Applications of hybrid genetic algorithms in

seismic tomography

Article in Journal of Applied Geophysics · November 2011

DOI: 10.1016/j.jappgeo.2011.08.005

CITATIONS READS

13 241

5 authors, including:

Pantelis Soupios Petros Mpogiatzis

Technological Educational Institute of Crete Harvard University
136 PUBLICATIONS 782 CITATIONS 16 PUBLICATIONS 22 CITATIONS

SEE PROFILE SEE PROFILE

Ahmet Tuğrul Başokur C. B. Papazachos

Ankara University Aristotle University of Thessaloniki
45 PUBLICATIONS 272 CITATIONS 203 PUBLICATIONS 3,313 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Global magnitude scaling relations, homogeneous earthquake catalogs and spatio-temporal analysis
of global intermediate depth & deep focus seismicity. View project

Solving Large Seismic Tomography Problems and Assessing the Uncertainty through the use of Direct
Sparse Methods, Combinatorics and Graph theory View project

All content following this page was uploaded by Irfan Akca on 23 December 2013.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
 Journal of Applied Geophysics 75 (2011) 479–489

Contents lists available at SciVerse ScienceDirect

Journal of Applied Geophysics

journal homepage: www.elsevier.com/locate/jappgeo

Applications of hybrid genetic algorithms in seismic tomography

Pantelis Soupios a,⁎, Irfan Akca b, 1, Petros Mpogiatzis c, 2, Ahmet T. Basokur b, 1, Constantinos Papazachos d, 3
a
Laboratory of Geophysics and Seismology, Department of Natural Resources and Environment, Technological Educational Institute of Crete, 3 Romanou Chalepa, 73133,
Chania, Crete, Greece
b
Ankara University, Faculty of Engineering, Department of Geophysical Engineering, Tandogan Kampusu, 06100 Ankara, Turkey
c
Harvard University, Seismology Group, Hoffman Laboratory, 20 Oxford Street, Cambridge, MA 02138, USA
d
Geophysical Laboratory, Department of Geology, Aristotle University of Thessaloniki, Greece

a r t i c l e i n f o a b s t r a c t

Article history: Almost all earth sciences inverse problems are nonlinear and involve a large number of unknown parameters,
Received 15 March 2011 making the application of analytical inversion methods quite restrictive. In practice, most analytical methods
Accepted 10 August 2011 are local in nature and rely on a linearized form of the problem equations, adopting an iterative procedure
Available online 24 August 2011
which typically employs partial derivatives in order to optimize the starting (initial) model by minimizing
a misfit (penalty) function. Unfortunately, especially for highly non-linear cases, the final model strongly de-
Keywords:
Geophysics
pends on the initial model, hence it is prone to solution-entrapment in local minima of the misfit function,
Genetic algorithms while the derivative calculation is often computationally inefficient and creates instabilities when numerical
Least-squares methods approximations are used. An alternative is to employ global techniques which do not rely on partial deriva-
Seismic tomography tives, are independent of the misfit form and are computationally robust. Such methods employ pseudo-ran-
domly generated models (sampling an appropriately selected section of the model space) which are assessed
in terms of their data-fit. A typical example is the class of methods known as genetic algorithms (GA), which
achieves the aforementioned approximation through model representation and manipulations, and has
attracted the attention of the earth sciences community during the last decade, with several applications al-
ready presented for several geophysical problems.
In this paper, we examine the efficiency of the combination of the typical regularized least-squares and ge-
netic methods for a typical seismic tomography problem. The proposed approach combines a local (LOM)
and a global (GOM) optimization method, in an attempt to overcome the limitations of each individual ap-
proach, such as local minima and slow convergence, respectively. The potential of both optimization methods
is tested and compared, both independently and jointly, using the several test models and synthetic refrac-
tion travel-time date sets that employ the same experimental geometry, wavelength and geometrical charac-
teristics of the model anomalies. Moreover, real data from a crosswell tomographic project for the subsurface
mapping of an ancient wall foundation are used for testing the efficiency of the proposed algorithm. The re-
sults show that the combined use of both methods can exploit the benefits of each approach, leading to im-
proved final models and producing realistic velocity models, without significantly increasing the required
computation time.
© 2011 Elsevier B.V. All rights reserved.

1. Introduction et al., 1998; Boschetti et al., 1996; Daley et al., 2004; Hole et al., 1992;
Lersviriyanantakul et al., 2006; Papazachos and Nolet, 1997; Vassallo
Traveltime tomography has been successfully employed at very dif- and Zollo, 2008; Weber, 2000). However, the practical application of to-
ferent spatial scales in order to recover the subsurface velocity model mographic methods is limited by the complex experimental geometry,
across various regions of interest (Ajo-Franklin et al., 2006; Bijwaard poor control of the available a priori model information and data errors.
Moreover, in certain environments such as Quaternary basins, the high
⁎ Corresponding author. Tel.: + 30 2821023037; fax: +30 2821023042. signal attenuation makes the collection of reliable and quantitatively
E-mail addresses: [email protected] (P. Soupios), sufficient traveltimes almost impossible, resulting in data gaps especial-
[email protected] (I. Akca), [email protected] ly from rays sampling low-velocity areas that are most often associated
(P. Mpogiatzis), [email protected] (A.T. Basokur), with higher attenuation.
[email protected] (C. Papazachos).
1
Tel.: +90 312 203 33 81; fax: +90 312 212 00 71.
Traveltime tomography traditionally involves the solution of a linear-
2
Tel.: +1 6174951172. ized system of the traveltime equations, using an appropriate initial
3
Tel.: +30 2310998510; fax: +30 2310998528. (starting) model. During the solution of a traveltime tomography

0926-9851/$ – see front matter © 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.jappgeo.2011.08.005
480 P. Soupios et al. / Journal of Applied Geophysics 75 (2011) 479–489

problem by using local optimization methods, the problem starts from an In the present work, the joint application of hybrid genetic algo-
initial velocity model and repeatedly “moves” to an improving neighbor- rithms and least-squares (representative of local and global optimiza-
ing set of solutions. This solution may be the different from the global op- tion, respectively) is examined and evaluated for seismic tomography
timum (determined using appropriate global optimization methods), applications. The robustness of both optimization methods has been
which is the optimal solution among all possible solutions. Unfortunately, tested and compared for the same or similar source–receiver geome-
the assumption of a linear behavior around the global minimum is only try and characteristics of the model structure (anomalies, etc.). A set
successful when adequate a priori information is available (good starting of seismic refraction synthetic (noise free) data is used for modeling.
model). Moreover, data errors, dependence of the ray geometry on the Specifically, cross-well, down-hole and typical refraction studies
model, poor model coverage and intrinsic correlation of available data using 24 geophones and five source shots were used to confirm the
usually render the problem non-unique and ill-conditioned. More specif- potential applicability of the genetic algorithms in seismic traveltime
ically, the main sources of non-uniqueness in traveltime tomography (e.g. tomography. Finally, a real dataset collected from a crosswell tomo-
Berryman, 1990) are, a) the model parameterization, b) the measurement graphic experiment was used for showing the applicability of the sug-
errors due to errors in source/receiver locations and incorrect first break gested hybrid tomographic method.
picking, and c) ghosts due to numerical instabilities and model perturba- In order to solve the forward modeling and estimate the traveltimes,
tion. For this reason the linearized system is often solved by appropriately a ray bending method was used. The root mean square (RMS) traveltime
adapted local optimization methods, such as constrained least-squares' or error is considered as the misfit function and calculated for the entire set
modified conjugate-gradient methods. Additional convergence problems of random velocity models of each generation. After each GA model gen-
can arise from the instabilities introduced by the calculation of the partial eration, the selection, crossover and mutation (typical process steps of
derivatives of the raypath (Jacobian) matrix required by the iterative im- genetic algorithms) operators are applied to create a new model gener-
provement of the starting (initial) velocity model using the linearized ation by evaluating the misfit of the individuals (velocity models). In
form of the problem (Aster et al., 2005; Sambridge, 1990). order to optimize the whole procedure computation time, the MATLAB
Since most geophysical inverse problems are nonlinear, hence have Distributed Computing Engine (MDCE) was incorporated using a multi-
non-linear misfit functions (e.g. rms travel-time error), the solution is core engine. During the tests, the initially fast convergence of the algo-
quite often trapped to local minima during the application of local opti- rithm (typically first five generations) is followed by progressively
mization methods. As a result, their success depends on the proximity of slower improvements of the reconstructed velocity models. Therefore,
the initial (starting) model to the “true” global-minimum solution. Some to improve the final tomographic models, a hybrid genetic algorithm
global optimization algorithms have been suggested to overcome this (GA) approach was adopted by combining the GAs with a local optimiza-
problem by sampling a wider search space in order to locate this global tion method applied after a few GA model generations, on the basis of
minimum (e.g. Boschetti and Moresi, 2001; Parker, 1999; Sambridge the convergence of the resulting models. This approach is shown to be ef-
and Mosegaard, 2002; Sen and Stoffa, 1995). Global techniques do not ficient, as it “guides” the solution search towards the neighborhood
rely on partial derivatives and the inversion of the Jacobian (raypath) model region of the global minimum.
matrix, hence they are computationally much more robust. In practice,
these methods start with a single or a set of best-guess or even randomly 2. Genetic algorithms
generated initial solutions, and use these solutions to sample “represen-
tative” sections of the model space. A typical example is the class of Ge- Genetic algorithms (GAs), which have been originally introduced by
netic algorithms, which progressively modify the initial solution set by John Holland (1975), belong to the group of stochastic global optimiza-
mimicking the evolutionary behavior of biological (genetic) systems, tion methods (randomized search methods) such as Monte Carlo and
until an acceptable result is achieved. simulated annealing. The GA concept is based on biological evolution,
Genetic algorithms (GAs) have been employed for the solution of seis- where the fittest model has higher probability of survival and reproduc-
mological (Nicknam et al., 2010; Nicknam and Eslamian, 2010; Wu et al., tion, while others (the weakest, lower-fitness value) have lower proba-
2008) and other geophysical and environmental problems (Akça and bilities. The main advantage of these methods is that they are able to
Basokur, 2010; Basokur and Akca, 2011; Basokur et al., 2007; Billings et define the global minimum, if it exists in the search space. However,
al., 1994; Buckles and Petry, 1992; Cavicchio, 1970; Currenti et al., an appropriate search space on the basis of the a priori information pro-
2005; Dal Moro, 2008; Dal Moro et al., 2007; Dal Moro and Pipan, 2008; vided by the user can be defined, allowing the efficient derivation of the
Davis, 1987, 1991; Goldberg, 1989; Jha et al., 2008; Jin and Madariaga, population of possible model solutions.
1993; Jishun et al., 2009; Krahenbuhl and Li, 2006; Louis et al., 1999; For the GA application, an initial population consisting of a fixed
Mathias et al., 1993; Rodriguez-Zuniga et al., 1997; Rucker and Ferre, number of individuals (velocity models in our case) is randomly gener-
2005; Sambridge, 1999a,b; Schwarzbach et al., 2005; Stoffa and Sen, ated. All physical parameters (genes) are combined as chromosomes to
1991). In GAs, a series of simulated genetic processes (namely coding, se- represent a state of the model. The collection of models constitutes a
lection, crossover and mutation) are employed along various generations population. After this step, the forward problem using ray tracing calcu-
to find the best models, based on their selected misfit values. This proce- lation is solved and the discrepancy between observed and calculated
dure (creating new generations) is terminated when a satisfactory fit, data (objective function) is estimated. The fitness is then used in the suc-
prescribed by the interpreter, is achieved. Moreover, due to the nature cessive selection and crossover process. The best individuals (i.e. the ve-
of genetic algorithms, the method offers the possibility of efficient locali- locity models with the smallest error) are selected to generate the next
zation of the most “promising” areas (global minimum) of the solution genes (offspring). The fitness of individuals is then used in the succes-
space. Since model derivatives are not involved, they are able to solve sive selection process. The winner of a tournament among a number
even very complex nonlinear problems, without the need of computing of randomly selected individuals (i.e. the velocity model with the smal-
numerically unstable gradient and/or curvature information. Moreover, lest error) has the right of passing its genes to the next generation. The
the linearization of the problem is not necessary since genetic algorithms genetic information exchanged between selected individuals is realized
are based only on direct model-space sampling. These factors make ge- by the crossover operator. During the mutation, selected bits of off-
netic algorithms particularly attractive for addressing complex real- springs based on the mutation probability, are flipped to allow for
world problems, which explains the fact that they are increasingly used small perturbations of the model parameters. In order to pass the fittest
for geophysical problems. However, the large dimensionality involved individual(s) of each generation to the next one, the elitism strategy is
in most geophysical optimization problems can reduce the efficiency of used avoiding possible loss of good individuals (Goldberg, 1989; Haupt
the genetic algorithm search, especially in terms of the required compu- and Haupt, 1998; Man et al., 1999). The iterations of this evolution the-
tational cost in order to obtain a good-quality solution. ory can be terminated after a fixed number of generations or when the
 P. Soupios et al. / Journal of Applied Geophysics 75 (2011) 479–489 481

fitness of an individual reaches a certain error threshold. A typical pseu- approach for the same problem, a variety of basic parameters (number
docode of a simple application of GA in seismic tomography using of generations and populations, variability of the model and number of
MATLAB (MDCE) is presented in Appendix A. nodes in the model), different initial velocity models (layered or homo-
geneous model) and several source–receivers configurations were used
3. Velocity reconstruction via genetic algorithms in order to generate and test different synthetic data sets. A typical exam-
ple is presented in Fig. 1, which shows the application of GAs in three dif-
The hybrid application of GAs and local optimization methods ferent crosswell (CW) or combined CW and vertical seismic profile (VSP)
presented here focuses on traveltime inversion for 2D velocity experimental geometries. The model consists of 16 nodes (nx =4,
models. Though in principle a priori information for the model nz =4) with a high velocity block (2.5 km/s) located in the center of
area can be used for the definition/limitation of the parameter the model, overlying a homogeneous (0.5 km/s) velocity medium, as
search space (especially for the GA approach), in practice a simple shown in Fig. 1. Synthetic traveltime data were generated for the preced-
layered velocity structure is often used as an initial model and ing synthetic velocity structure by performing forward calculations. The
search limits of the parameter search are specified in relation to first test consists of three sources and seven receivers, resulting in 21
this starting velocity model. The initial population for the GA algo- rays and a poor raypath coverage of the upper and lower part of the
rithm is generated using random distributions of model parameters model. The second model consists of two sources and eight receivers,
over the whole search space. Models are coded using a binary resulting in 16 rays slightly improving the raypath coverage. The last
scheme in order to facilitate their processing with genetic operators model consists of four sources and eight receivers, resulting in 32 rays
and was subsequently used for all GA “experiments”. Once the cod- and corresponds to an improved raypath coverage test, in comparison
ing scheme was selected, all the encoded model parameters are to the two other models. Using the GA approach, the optimum velocity
merged into a long bit-string (chromosome), which is modified model (in both shape and amplitude) is reconstructed for the best ray-
(updated) by the algorithm. This approach means that higher pa- path coverage (third model) after roughly 45 generations.
rameter resolution requires relatively long parameter strings. The Since the high velocity anomalies are much better reconstructed
algorithm then determines the fitness of the individual models by than low velocity anomalies in all seismic tomographic problems (re-
decoding each bit-string “binary” information into the velocity gardless of the experimental geometry), the application of the proposed
model parameters. Selection, crossover and mutation operators act algorithm was tested for more complex velocity structure problems.
on individuals of previous generations. Generally, the number of Fig. 2 shows the application of the proposed algorithm in a typical CW
better fitted models is expected to increase in the population model consisting of 24 nodes (unknowns), with a homogeneous veloc-
throughout the evolution process (Sen and Stoffa, 1995; Soule and ity background and two anomalies (one positive and one negative) in
Foster, 1998). Selection of individuals for mating is based on a “tour- the upper and lower part of the model, respectively. The test on the
nament selection” procedure, which corresponds to selecting the left (Fig. 2A) incorporates 51 sources (left side, red filled circles) and
winner of a tournament among a certain number of randomly cho- 51 receivers (right side, green filled circles), resulting in 2601 raypaths,
sen members of the population based on misfit values. The winners already tested by Soupios (2000). The second model (Fig. 2B) consists of
of two sequential tournaments become the mates selected for cross- a much smaller number of sources (5) and receivers (9), resulting in
over. Once the mates are chosen, a “scattered crossover” takes place only 45 raypaths. The main concept behind this test was to examine
between the genes of two individuals. Since the crossover probabil- whether the proposed algorithm can efficiently reconstruct velocity
ity acts in a different manner for each application/problem (Akca, models from much smaller data sets.
2010; Basokur and Akca, 2011), after several trials the value of Fig. 2C and D presents the final tomographic velocity models of Fig. 2A
65% for all experiments was selected and used as the optimum for and B, respectively. The tomographic velocity model of Fig. 2C was calcu-
our application. A random number chosen uniformly within the [0 lated after the inversion of first traveltime arrivals using a standard least-
1] range designates which mate contributes for that bit of the off- squares method and applying appropriate damping and smoothing regu-
spring (if the value of the random number is greater than 0.65 cross- larization. Both tomographic models were reconstructed by more than
over does not occur, else a second mate provides the gene). Random 85% in both amplitude and shape of the velocity anomalies used. The
mutations with a probability of 0.2% are permitted, if the micropo- RMS traveltime error of the first (Fig. 2C) and the second (Fig. 2D) velocity
pulation option is not turned on. Notice that micropopulation redis- models were ~2.1% and ~2.3%, respectively. It is interesting to notice that
tributes the models over the search space (keeping the best-fit GAs managed to reconstruct a high quality velocity structure even using
model) when the model population becomes nearly homogenous. 58 times less raypaths (45 raypaths instead of the 2601 raypaths) than
The crossover and mutation probabilities were fixed to these values the model of Fig. 2A. Computationally, the model used in the case of
after several trials for the examined problem, as previously men- Fig. 2A (121 nodes) for 2601 raypaths and a local optimization method,
tioned. It should be pointed out that for a different inverse problem, required an average CPU time of 25 min (on an Opteron 8220 2.8Ghz pro-
or for different settings of the population size and number of param- cessor) to solve the forward and inverse problem (Fig. 2C). On the other
eters a different setting of crossover probability may yield the best hand, the model used in Fig. 2B (24 nodes) for 45 raypaths and with the
result. application of GAs (30 generations) required on the same CPU an average
The solution of the forward modeling involves the accurate calculation time of 2150 min (36 h) to reconstruct the velocity model as shown in
of the first arrivals for a given velocity model. Rays for the first arrivals Fig. 2D. This result suggests that the simple application of GAs is practical-
were calculated using a two-step procedure. In the first step, graph theory ly prohibited as the data and model size increase, even if high CPU power
and the revised algorithm of Dijkstra (Moser, 1991) were used for esti- is available, a feature which typically characterizes global optimization
mating the traveltime of the wave front at each node of the grid (for a methods in comparison to local optimization techniques.
given receiver) and provide preliminary ray-paths. These paths were In general, GAs provide a fast initial convergence followed by pro-
further optimized in the second step by a revised bending algorithm, gressively slower improvements (Boschetti, 1995a, b). Thus, the form
which has been shown to have superior convergence properties (Moser of the misfit curve variation against GA iteration suggests that the algo-
et al., 1992; Thurber and Ellsworth, 1980). This combined approach rithm should be stopped when an approximate solution has been
ensures that the final ray corresponds to a global and not local traveltime found, since further improvements may be much more time consuming.
minimum, as the graph theory selects an initial ray-path close to the Hence, it is quite clear that significant improvements of the GA perfor-
global minimum (Mpogiatzis, 2006; 2010; Mpogiatzis et al., 2009). mance may be obtained by combining the genetic algorithm with a
In order to evaluate and test the efficiency of the proposed algorithm local optimization method, in order to direct the search to the region
and to compare it with the application of typical local optimization close to the global solution.
482 P. Soupios et al. / Journal of Applied Geophysics 75 (2011) 479–489

Fig. 1. Tomographic results derived from the application of genetic algorithm for three different source–receiver geometries using the same velocity model. Models A1, A2 and A3
depict the ideal (synthetic) anomaly for different experimental source–receiver geometry and raypath coverage. The corresponding models B1, B2 and B3 were produced using a
standard GA approach after 30 generations. Notice that model C3 is similar to B3 but after 45 generations.

3.1. Hybrid genetic algorithms — combination of global and local 0.8 ms), showing a stagnation of the GA process (almost constant
optimization misfit), the velocity models with a misfit less than the average mis-
fit (best 50% of the population) are selected and inserted as initial
Considering their individual characteristics, the combination of global models in a standard least-squares inversion. The updated (im-
and local optimization can provide an efficient way for solving nonlinear proved) velocity models after the inversion are re-transformed
inverse problems. Cary and Chapman (1988) performed waveform inver- into binary format in order to be inserted back into the original
sion by the combined use of the Monte-Carlo and gradient methods, GA population. In other words, we employ the least-squares meth-
while Stork and Kusuma (1992) combined genetic algorithm with the od to improve a part of the GA population, in an attempt to speed
steepest descent method for solving the residual static correction forward up the convergence, maintaining the main benefits of the global op-
problem. Mendes (2009) presented a hybrid algorithm, combining a timization approach. The selection, crossover and mutation are then
Monte-Carlo approach with the SIRT method for first arrival tomography. applied for the creation of the next GA generation.
Chunduru et al. (1997) systematically described seven hybrid optimiza- • Following previous suggestions (e.g. Boshetti, 1995a), if the best fit
tion methods combining very fast simulated annealing and conjugate model from this hybrid GA approach has an RMS error less than a
gradient (CG) for 2D imaging of electrical resistivity and seismic data. prescribed minimum level (e.g. 0.5 ms), we use this best-fit velocity
As a result, the combined use of global and local optimization, exploiting model from the population as the starting velocity model for a final
the advantages of each method, has created a new trend in inversion least-squares inversion, in order to determine the optimum velocity
problem research. model. In other words, we use least squares when we believe that
In the present work we present an efficient and robust method for we are already close enough to the global minimum to overcome
nonlinear travel-time inversion by combining the standard regularized the convergence problem of the GA algorithm.
least-squares solution and genetic algorithms. Specifically, we consid-
ered a hybrid genetic algorithm, where the GA algorithm is applied
It is clear that the selection of the optimum parameters for the appli-
and the least-squares approach is inserted in the process in two stages:
cation of the hybrid GA approach (i.e. minimum number of iterations
• After a minimum number of GA generations (typically 5) if the mis- and misfit threshold to start improving the best 50% of the GA popula-
fit change among generations is less than prescribed level (e.g. tion, model minimum misfit to apply a final least-squares inversion)
 P. Soupios et al. / Journal of Applied Geophysics 75 (2011) 479–489 483

Fig. 2. Two crosswell (CW) experiments with different experimental (source–receiver) geometries, homogeneous velocity background and two anomalies (one positive and one
negative) in the upper and lower part of the model, respectively. A) A CW experiment with 51 sources (red circles) and 51 receivers (green polygons) (2601 raypaths) consisting
of 121 nodes, B) a similar model consisting of 5 sources and 9 receivers, resulting in 45 raypaths using a coarser grid of 24 nodes, C) final tomographic velocity model as extracted
from the application of local optimization (least-squares) method is depicted. The reconstruction of the ideal low and high velocity anomalies are shown. D) Final velocity model
and the reconstructed high and low velocity anomalies determined from the application of a standard genetic algorithms approach (global optimization methods). (For interpre-
tation of the references to color in this figure legend, the reader is referred to the web version of this article.)

previously described is empirical and should be calibrated using several however the micropopulation perturbation is also used. This additional
tests on synthetic and real data. model perturbation is applied when the average and the best misfit
In order to test the applicability and efficiency of the proposed hybrid are very similar, suggesting a very limited variability of the population
algorithm, we employed a typical dipping layered model with increasing models; in this case the population is randomly perturbed in order to
velocity with depth. A CW and VSP measurement configuration was con- allow the algorithm to better “explore” the model space. As is depicted
sidered with three sources (at 7, 21 and 35 m depth) and seven receivers in Fig. 3E, the final misfit is about ~1.3 ms (higher than the other
on the surface at 5 m and at several depths (at 7, 15, 21, 28, 35 and 42 m), models) but the finally reconstructed velocity model after 30 genera-
resulting at 21 raypaths (Fig. 3C — upper figure), while the model consists tions, is much closer to the ideal velocity model (Fig. 3C — upper figure),
of 21 nodes (nx=3, nz =7). The starting model for all tests was a homo- with the dipping interface much better delineated.
geneous (1.0 km/s) velocity model. Different population sizes were used Model B of Fig. 3 depicts the reconstructed velocity model by ap-
and presented for the standard GA approach (Fig. 3B), as well as the hy- plying a standard GA algorithm after 70 generations, with a final mis-
brid GA method (Fig. 3A, D and E). The final velocity models can be com- fit of ~ 0.8 ms. The effect of inserting the least-squares acceleration in
pared with the input velocity model (Fig. 3C — upper figure), while the the algorithm is evident. In all examined cases (Fig. 3, models A, D and
tomographic velocity model from the application of a standard least- E) after 5 generations the least-squares inversion is turned on and
squares method (8 iterations and a misfit of 0.29 ms — RMS: 4.16%) is significantly reduces the traveltime misfit. As a result, after 15 gener-
also shown for comparison (Fig. 3C — lower figure). The misfit curve for ations the algorithms provide fairly reliable velocity models. On the
all GA tests is also presented, in order to assess the convergence rate of contrary, in order to obtain a similar (in quality) velocity model
each approach, as well as to indirectly assess the reliability of the final from the standard GA approach (model B), more than 30 generations
resulting velocity models. need to be applied.
Models A and D of Fig. 3 correspond to the application of the hy- It is also interesting to compare the results with the final tomograph-
brid algorithm by applying different population sizes, equal to 15 ic velocity model (Fig. 3C — lower model) obtained with a standard reg-
and 5 times the number of velocity nodes, respectively. It is clear ularized least-squares method (8 non-linear iterations). It is clear that
that the insertion of the least-squares inversion in the GA process sig- this model is quite different from the real (input) model, most probably
nificantly speeds up convergence, without significantly affecting the due to the very poor data-coverage of the specific synthetic example
global optimization process. Moreover, the final misfit for both tests (only 21 raypaths used to determine a 21-node model). This discrepan-
after 30 generations is similar (~0.7 ms). This observation suggests cy, as well as the previous results clearly show that, a) the proposed hy-
that the estimated velocity model does not critically depend on the brid algorithm (combination of GA with a least-squares method) is
population size; hence usage of relatively fewer models (population working efficiently, mainly accelerating the solution convergence and,
size) in the GA algorithm does not affect the reliability of the final b) is more efficient in reconstructing tomographic velocity models
model, due to the insertion of the least-squares step, contrary to than the standard least-squares method, especially when a rather limit-
what is usually observed in the standard GA algorithm. ed data-set is available.
Model E depicts the application of the hybrid algorithm with a typical The main disadvantage of the GAs approach in seismic tomography
population size (10 times the number of nodes-velocity parameters), applications is the CPU time required for the estimation of the final
484 P. Soupios et al. / Journal of Applied Geophysics 75 (2011) 479–489

Fig. 3. A dipping layered model used in standard and hybrid GA inversion with different numbers of population. 3 sources and 7 receivers are used for the synthetics (21 raypaths),
while the model grid consists of 21 velocity nodes. The starting model for all applications is a homogeneous velocity model (1 km/s). Models A and D use the hybrid algorithm with
different population sizes (5 and 15, respectively), while model B corresponds to the standard application of GA. Model E corresponds to the application of the hybrid algorithm,
with the additional use of a micropopulation model perturbation when the misfit is almost invariable. The upper model C presents the synthetic velocity model and the ray cov-
erage, while the lower tomographic model is derived by a standard least-squares (local optimization) method. For all GA models (A, B, D and E) the misfit variation curve is also
presented. The red and black curves present the best-fit and average model of each generation, respectively. (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of this article.)

velocity model. To overcome this problem, the hybrid algorithm was Synthetic traveltimes were estimated and the corresponding raypaths
modified for multi-core engines using the MDCE module of MATLAB. are also presented in Fig. 4B.
Specifically, a job manager was defined and the idle workers were used In Fig. 4, the obtained velocity models after the application of the pro-
for running the hybrid algorithm (see Appendix A) in a pseudo-parallel posed hybrid GA approach for several generations (gen: 1, 10, 20, 28) are
way, distributing the population to the available workers. This approach presented. The final tomographic velocity model after 14th iterations
allows the application of the proposed approach in more complicated (Fig. 4A) by applying a least-squares method is also shown. Notice that al-
traveltime tomography problems involving larger datasets. though the best-fit velocity model acquired after 28 generations from the
Taking advantage of the improvement of the computing time using application of hGALSA has a misfit of ~2 ms (Fig. 4C) and the correspond-
the MDCE module of MATLAB, an additional test was performed in ing optimal tomographic velocity model after the application of the least-
order to check the applicability and reliability of the proposed hybrid squares algorithm has a smaller misfit (~0.75 ms, RMS=4.08%, Fig. 4A),
GA/least-squares algorithm (hGALSA). A typical refraction experiment the synthetic model is much better reconstructed (in both amplitude
configuration was examined using 5 sources (at 0, 50, 175, 300 and and shape) by the hGALSA than the LOM.
320 m) along the ground surface, sampled at 24 geophones with a spac-
ing of 10 m (g1: 60 m … g24: 290 m). In total, 120 raypaths were used 4. Application to real crosswell data
to define 45 unknown model parameters (9 by 5 model velocity nodes).
A test model including a layered velocity model with the addition of a Following the experiments on synthetic data, the proposed algorithm
low velocity body (e.g. soft-sediment ditch) was considered (Fig. 4B). was tested on a real crosswell experiment. The data were collected from
 P. Soupios et al. / Journal of Applied Geophysics 75 (2011) 479–489 485

Fig. 4. Inverted velocity models derived for a typical seismic refraction experiment, as calculated from the application of the hybrid GAs (hGA) and the least-squares (LOM). The
velocity model represents the synthetic velocity structure, where the source–receiver geometry and corresponding raypaths are also shown. The remaining models presented
(C, D, E and F) show the velocity model after several generations during the application of the hybrid GA approach. The velocity model on the top (A) presents the final tomographic
models after several iterations (14th) of the least-squares' methods.

a seismic survey experiment that was conducted at Thessaloniki city in at the area of the survey. The dense urban environment made the appli-
Northern Greece, in the framework of the construction of the new met- cation of large/full-scale surface geophysical prospecting impossible,
ropolitan underground railway of the city. The initially designed railway thus it was decided that the optimum approach was the drilling of bore-
root crosses Byzantine walls and other monuments. The main purpose of holes at either side of the wall in order to perform joint cross-hole/down-
this survey was the exact and rapid determination of the extent of the hole tomography measurements. The distance between the boreholes
foundations of a section of the Byzantine wall prior to archeological exca- was about 8 m. Unfortunately, during the casing procedure of the first
vations and construction works. The specific wall remains were visible (left) borehole excessive cement grout usage was required (probably
on the surface, with an approximate height of about 1 m from the surface due to a subsurface cavity). As a result a wide high velocity area was
486 P. Soupios et al. / Journal of Applied Geophysics 75 (2011) 479–489

expected to be identified at the bottom left section of the study area. On methods fail to reconstruct the real velocity model since the solution is
the surface, a total of 12 P-wave geophones were installed between both trapped to local minima, either due to some bad data (low signal to
boreholes, the first with offset 0.8 m from the first borehole and the rest noise ratio-poor quality signals) or because the starting model is close
every 0.6 m. Furthermore, 3 borehole triaxial geophones (GEOSTUFF- to some basin of attraction that correspond to a local minimum different
BHG-2) with 1 m spacing were used and by their sequential placement than the “real” solution. The final hybrid GA velocity model (Fig. 5B) is fa-
inside the first borehole, recordings up to depth of about 13 m from vorably comparable to the final velocity model of Mpogiatzis (2010),
the surface were performed. P-waves were generated from a borehole which was estimated by a joint-inversion of both P- and S-wave record-
source at the second borehole which placed at different depths with a ings and the application of different data weighting (according to their
spacing of 1 m. Moreover, a hammer surface source was also used to pro- quality), as well as a finer model grid (Fig. 5E).
vide additional surface-to-borehole recordings.
The previous measurement setup yielded a satisfying coverage of
the investigation area by seismic raypaths (Fig. 5A). A total of 24 sources 5. Conclusions
were used from which high resolution traveltimes were collected from
25 receivers (both on the surface and in the borehole). The final data set The results presented in Figs. 3 and 4, provide evidence that the pro-
consists of 341 first breaks of P-waves after removing waveforms with posed hybrid GA/least-squares algorithm produces improved results
poor signal quality. The tomographic model consisted of 77 velocity (better velocity model reconstruction regarding both amplitude and
nodes (7 nodes in horizontal, 11 nodes in the vertical direction, with a shape) that the standard least-squares method for seismic tomography.
node spacing of ~1.5 m in both directions). Using the borehole logs, Moreover, the proposed hybrid algorithm combines the benefits of a
an initial velocity model with a positive depth-gradient was used, con- global and a local optimization method and is implemented in MATLAB
sisting of 3 layers (0.5, 1.2 and 2.0 km/s). for multicore engines in order to solve large dimensional problems and
The final velocity model after the application of the hybrid GA algo- achieve fast location of global minima, providing robust and reliable to-
rithm for 30 generations is presented in Fig. 5B, exhibiting a final misfit mographic images. It should be also noted that, the hybrid GA can better
of ~3.2 ms. Moreover, a comparison of this model with the final inverted reconstruct the low velocity model areas than the traditional least-
velocity models by using different local optimization methods (Occam squares method approach. The final velocity model obtained by this al-
method minimizing the 2nd derivative — Fig. 5G, smoothness constraint gorithm is practically independent of the initial model, while the algo-
minimizing the 2nd derivative — Fig. 5F, classical Levenberg–Marquardt rithm requires relatively fewer data in order to obtain realistic and
regression with damping — Fig. 5H) with the same configuration (77 un- reliable models, an advantage which can be quite important in real
known parameters) is also shown. Seems that the local optimization world study cases.

Fig. 5. Final velocity models for the Byzantine wall foundation study area examined in this work. (A) The sources–receivers configuration as well as the raypath coverage for the real
experiment is presented to highlight the good coverage of the study area. (B, C, D) The reconstructed velocity models by using the hGA method after 30, 20 and 5 generations, re-
spectively, are shown. (F, G, H) The final tomographic velocity models by applying different local optimization methods (smoothness constraint, Occam inversion and Levenberg–
Marquardt with damping, respectively) to the same experimental configuration are given. (E) The final tomographic velocity model (Mpogiatzis, 2010) adopting different weights
to data, topography corrections, threshold to “bad” measurements and a finer grid is shown. The solid black line depicts the relics of the ancient Byzantine wall.
 P. Soupios et al. / Journal of Applied Geophysics 75 (2011) 479–489 487

Acknowledgments ul = groupsize; % upper limit

NumberofTasks = npopsiz/groupsize;
The authors would like to thank the Seismological Station of the for itask = 1:NumberofTasks
Aristotle University of Thessaloniki, for providing computing time in
createTask (job1,@refraction,1,{igen,groupsize,xmin,zmin,dx,dz,
distributed machines. PS would also like to thank the Greek Ministry
xmax,zmax,parent (dl:ul,:)});
of Education for supporting a 6 month visit to Ankara University, as
dl = dl + groupsize;
well as Prof. F. Boschetti for providing valuable information, codes
and help for the completion of this project. We would also like to ul = ul + groupsize;
thank the Assoc. editor, Dr. G. Dal Moro and an anonymous reviewer end
for their constructive comments and suggestions which helped to submit (job1)
improve the manuscript. waitForState (job1,'finished');
out = getAllOutputArguments (job1);
ttminback = cell2mat (out (:,1)')';
Appendix A. A template of the application of GAs in
seismic tomography finished_jobs = findJob (jm,'State','finished');
destroy (finished_jobs)
% Input geometrical parameters of 2D model and initial model end
velocities: % Compute measured-calculated traveltimes misfit by means of RMS
% X-limits of model: xmin, xmax load measured_tt.dat;
% Node spacing in x and z directions :dx, dz misfit = norm (measured_tt-ttminback)/sqrt (length(ttminback));
% Max depth of model: zmax average_misfit (igen) = sum (misfit)/npopsiz;
% Initial model velocities: v [bestfit,loc] = min (misfit); %best fitted individual
fid = fopen('InitModel.md','r'); fittestvel = parrent (loc,:);
[xmin, xmax, dx, dz, zmax,v] = finput(fid); % Is just a five-points filtering along the misfit curve to be able to
xnodes = ((xmax-xmin)/dx) + 1;%Number of nodes along x- % check the changes more safely.
direction if igen N 4;
znodes = ((zmax-zmin)/dz) + 1;%Number of nodes along z- iav = igen-2;
direction bf_av (iav) = average_misfit (iav)*0.40+ …
nparam = xnodes*znodes;%Number of unknown parameters average_misfit (iav-1)*0.20+ …
%Define parameter search spaces for GA by giving a %15 variance average_misfit (iav + 1)*0.20+ …
to initial average_misfit (iav-2)*0.10+ …
%values average_misfit (iav + 2)*0.10;
parmin(1:nparam) = v/1.15; end;
parmax(1:nparam) = v*1.15; % HYBRID GA
%GA parameter definitions % Based on Boschetti's PhD, after the fast convergence that GA
microaga = 1;%use of micropopulation instead of mutation % appear, no further improvement are expected and we should
pcross = 0.65;%crossover probability maybe
pmutate = 0.02;%mutation probability if (microga turned off) % apply local search methods for reaching the final optimum
maxgen = 30;%number of generations solution
npopsiz = 10*nparam;%An empirical value for the population size if bestfit b 0.5
%Generate the initial population randomly % Insert the fittest velmodel as initial model for inversion
[iparent] = initial(nparam,parmin,parmax,npopsiz); niter = 10;%number of iterations
%Input measured data (traveltimes) [RMS,mnew] = INVgen (xmin,zmin,dx,dz,xmax,zmax,fittest-
load SR.dat vel,idealtt,niter);
%Main evolutionary loop if RMS b bestfit
for igen = 1:maxgen parent (loc,:) = mnew;
% Decode the binary coded strings into decimal end
parents = decode(iparent); end
if npopsiz N 100 if (igen N 5)
% run forward calculations on local machine misfit_change = abs (bf_av (iav-1)-bf_av (iav));
ttminback = refraction(igen,xmin,zmin,dx,dz,xmax,zmax, end
parents);%vectorized % If the misfit change between generations is less than 10%
else if (igen N 5) && (misfit_change b 0.1)
% distribute the calculations over the cpu's % define the fittest velocity models compared to the average misfit
jm = findResource('scheduler','type','jobmanager','LookupURL',' [ix] = find (misfit b (average_misfit (igen)));
HOSTNAME'); nmoditer = length (ix);
job1 = createJob (jm); for ii = 1:nmoditer;
% SR.dat : Sources - Receivers location % Run the forward modeling once more for each selected vel model
% refraction.m : an algorithm to solve the forward problem % to estimate the Jacobian (derivative) matrix
set (job1,'FileDependencies',{'refraction.m','SR.dat'}); bestvel = parent (ix(ii),:);
dl = 1; % down limit ttminback = refraction (igen,1,xmin,zmin,dx,dz,xmax,zmax,
groupsize = npopsiz/(jm.NumberOfIdleWorkers); bestvel);
488 P. Soupios et al. / Journal of Applied Geophysics 75 (2011) 479–489

Berryman, J.G., 1990. Stable iterative reconstruction algorithm for nonlinear traveltime
% Insert the following velmodels as initial model for inversion
tomography. Inverse Problems 6, 21–42.
[RMS,mnew] = INVgen (igen,xmin,zmin,dx,dz,xmax,zmax, Bijwaard, H., Spakman, W., Engdahl, E.R., 1998. Closing the gap between regional and
bestvel,idealtt,niter); global travel time tomography. Journal of Geophysical Research 103, 30055–30078.
Billings, S., Kennett, B., Sambridge, M., 1994. Hypocentre location: genetic algorithms
% In order to include the inverted vel model to the initial polula- incorporating problem specific information. Geophysics Journal International
tion we 118, 693–706.
Boschetti, F., 1995b. A staged genetic algorithm for tomographic inversion of seismic
% should keep the resulted vel models into the parmin - parmax refraction data. Expl. Geophys. Vol. 25, 173–178.
limits Boschetti, F., Moresi, L., 2001. Interactive inversion in geosciences. Geophysics 66 (4),
1226–1234.
new_min = max (mnew);
Boschetti, F., Dentith, M.C., List, R.D., 1996. Inversion of seismic refraction data using
new_max = min (mnew); genetic algorithms. Geophysics 61 (6), 1715–1727.
save ('parmin.mat','new_max','new_min'); Boshetti, F., 1995a. Application of Genetic Algorithms to the Inversion of Geophysical
Data. PhD Thesis in Mathematical Geophysics, University of Western Australia.
% The MNEW is the updated velocity model after Inversion and Buckles, B.P., Petry, F.E., 1992. Genetic Algorithms: Introduction and Applications. IEEE
transformed to Computer Society Press, Catalog #2935, Los Alamitos, CA.
Cary, P.W., Chapman, C.H., 1988. Automatic 1D waveform iversion of marine seismic
% binary format to include these best vel models to the original refraction data. Geophysical Journal 93, 527–546.
population Cavicchio, D.J., 1970. Adaptive Search Using Simulated Evolution. Unpublished doctoral
dissertation, University of Michigan, Ann Arbor, MI.
if RMS b misfit (ix (ii)) Chunduru, Raghu K., Sen, Mrinal K., Stoffa, Paul L., 1997. Hybrid optimization for geo-
binarray = code2 (nparam,mnew,ig2,g0,g1); physical inversion. Geophysics 62 (4), 1196–1207.
Currenti, G., Del Negro, C., Nunnari, G., 2005. Inverse modelling of volcanomagnetic
iparent(:,ix (ii)) = binarray;
fields using a genetic algorithm technique. Geophysical Journal International 163,
misfit (ix (ii)) = RMS; 403–418.
end Dal Moro, G., 2008. VS and VP vertical profiling via joint inversion of Rayleigh waves
and refraction travel times by means of bi-objective evolutionary algorithm. Jour-
end; % End of the inversion of all the fittest models nal of Applied Geophysics 66, 15–24.
end; Dal Moro, G., Pipan, M., 2008. Joint inversion of seismic data by means of multi-
objective evolutionary algorithms. Society of Petroleum Engineers - 70th European
% End of the check about the partial inversion before continue the Association of Geoscientists and Engineers Conference and Exhibition - Incorporat-
gen ing SPE EUROPEC 2008, Vol. 2, pp. 1053–1057.
Dal Moro, G., Pipan, M., Gabrielli, P., 2007. Rayleigh wave dispersion curve inversion via
% Selection crossover and mutation
genetic algorithms and Marginal Posterior Probability Density estimation. Journal
mate1 = selecter (misfit,npopsiz); of Applied Geophysics 61, 39–55.
mate2 = selecter (misfit,npopsiz); Daley, T.M., Majer, E.L., Peterson, J.E., 2004. Crosswell seismic imaging in a contaminat-
ed basalt aquifer. Geophysics 69 (1), 16–24.
ichild = iparent(n,mate1); Davis, L., 1987. Genetic Algorithms and Simulated Annealing. Morgan Kaufmann Pub-
ras = rand(1,nchrome); lishers, Inc.
Davis, L., 1991. Handbook on Genetic Algorithms. Van Nostrand Reinhold.
mer = find (ras b =pcross); Goldberg, D.E., 1989. Genetic Algorithms in Search, Optimization, and Machine Learn-
ichild (mer,1:npopsiz) = iparent (mer,mate2); ing. Addison-Wesley Publ. Co., Inc.
Haupt, R., Haupt, S., 1998. Practical Genetic Algorithms. John Wiley and sons.
% end of evolution Hole, J.A., Clowes, R.M., Ellis, R.M., 1992. Interface inversion using broadside seismic re-
% code the new generation fraction data and three-dimensional travel time calculations. Journal of Geophysi-
cal Research Vol. 97 (No. B3), 3417–3429.
iparent = newgen (ielite,npossum,ig2sum,ibest,npopsiz,ichild,
Holland, J., 1975. Adaptation in Natural and Artificial Systems. University of Michigan
iparent); Press.
if microga==1; Jha, M.K., Kumar, S., Chowdhury, A., 2008. Vertical electrical sounding survey and resis-
tivity inversion using genetic algorithm optimization technique. Journal of Hydrol-
[iparent] = gamicro (npopsiz,nchrome,iparent,ibest); ogy 359, 71–87.
end Jin, S., Madariaga, R., 1993. Background velocity inversion with a genetic algorithm.
Geophysical Research Letters 20, 93–96.
if best b 1e-09 Jishun, P., Xinjian, W., Xiankang, Zh., Zhaofan, Xu., Ping, Zh., Xiaofeng, T., Suzhen, P., 2009.
break 2D multi-scale hybrid optimization method for geophysical inversion and its applica-
tion. Applied Geophysics 6 (4), 337–348. doi:10.1007/s11770-009-0034-x.
end Krahenbuhl, R.A., Li, Y., 2006. Inversion of gravity data using a binary formulation. Geo-
% save results physical Journal International 167, 543–556.
string_gen = int2str (igen); Lersviriyanantakul, Ch., Aiyarak, P., Lohawijarn, W., 2006. An application of genetic al-
gorithms in seismic refraction interpretation. Walailak Journal Sciences & Technol-
save_file = strcat('Results',string_gen,'.mat'); ogy 3 (2), 145–165.
save (save_file, 'parents', 'xp', 'zp', 'bf', 'average_misfit','misfit'); Louis, S.J., Chen, Q., Pullammanappallil, S., 1999. Seismic velocity inversion with genetic
algorithms. Washington D.C., July 6–9, 1998 CEC99, 1999 Congress on Evolutionary
end %%%% Close the iteration over all generations Computation, Mayflower Hotel, pp. 855–861.
fclose('all'); Man, K.F., Tang, K.S., Kwong, S., Halang, W.A., 1999. Genetic Algorithms: Concepts and
Designs. Springer-Verlag, London.
%END OF CODE Mathias, K., Whitley, D., Stork, C., Kusuma, T., 1993. Staged hybrid genetic algorithm
search for seismic data imaging. Presented at the Internat. Conf. on Evolutionary
References Computing, ICEC 1994.
Mendes, M., 2009. A hybrid fast algorithm for first arrivals tomography. Geophysical
Ajo-Franklin, J.B., Urban, J.A., Harris, J.M., 2006. Using resolution-constrained adaptive Prospecting 2009 (57), 803–809. doi:10.1111/j.1365-2478.2008.00755.x.
meshes for traveltime tomography. Journal of Seismic Exploration 14, 371–392. Moser, T.J., 1991. Shortest path calculation of seismic rays. Geophysics 56, 59–67.
Akca I., 2010. Melez genetic algoritmalar ile yapi-tabanli model gercekleme. PhD The- Moser, T.J., van Eck, T., Nolet, G., 1992. Hypocenter determination in strongly heteroge-
sis, 94 p (Abstract in english). neous earth models using the shortest path method. Journal of Geophysical Re-
Akça, I., Basokur, A.T., 2010. Extraction of structure-based geoelectric models by hybrid search 97, 6563–6572.
genetic algorithms Jan.-Feb. 2010 Geophysics Vol. 75 (No 1), F15–F22. Mpogiatzis, P., 2006. Two dimensional inversion of seismic refraction data. MSc Thesis,
doi:10.1190/1.3273851. Department of Geology, Aristotle University of Thessaloniki, 164 p. (in greek).
Aster, R.C., Thurber, C.H., Borchers, B., 2005. . 2005 Parameter estimation and inverse Mpogiatzis, P., 2010. Combinatorial seismic tomography. PhD Thesis, Aristotle Univer-
problems. Academic Press, p. 301. sity of Thessaloniki, Greece, pp. 220 (Abstract in englsih).
Basokur, A.T., Akca, I., 2011. Object-based model verification by a genetic algorithm ap- Mpogiatzis, P., Papazachos, C., Tsourlos, P., Vargemezis, G., 2009. Joint inversion of P
proach: application in archaeological targets. Journal of Applied Geophysics. and S traveltime tomography data using Poisson ratio and cross-gradient con-
doi:10.1016/j.jappgeo.2011.05.004. straints. EAGE, Near Surface 2009.
Basokur, A.T., Akca, I., Siyam, N.W.A., 2007. Hybrid genetic algorithms in view of the Nicknam, A., Eslamian, Y., 2010. A hybrid method for simulating near-source, broad-
evolution theories with application for the electrical sounding method. Geophysi- band seismograms: application to the 2003 Bam earthquake (Mw 6.5). Tectono-
cal Prospecting 55, 393–406. physics. doi:10.1016/j.tecto.2010.03.007.
 P. Soupios et al. / Journal of Applied Geophysics 75 (2011) 479–489 489

Nicknam, A., Abbasnia, R., Bozorgnasab, M., Eslamian, Y., 2010. Synthesizing strong mo- Sen, M.K., Stoffa, P.L., 1995. Global Optimization Methods in Geophysical Inversion.
tion using empirical Green's function and genetic algorithm approach. Journal of Elsevier Science Publications, Netherlands.
Earthquake Engineering 14 (4), 512–526. Soule, T., Foster, J.A., 1998. Effects of code growth and parsimony pressure on popula-
Papazachos, C., Nolet, G., 1997. P and S deep velocity structure of the Hellenic area tions in genetic programming. Evolutionary Computation 6 (4), 293–309.
obtained by robust non-linear inversion of travel times. Journal of Geophysical Re- doi:10.1162/evco.1998.6.4.293.
search 102, 8349–8367. Soupios, P.M., 2000. NonLinear Inversion - Tomography of Seismics. PhD Thesis, De-
Parker, P.B., 1999. Genetic Algorithms and Their Use in Geophysical Problems. Ph.D. partment of Geology, Aristotle University of Thessaloniki, Greece, p. 211 (in greek).
Thesis, Ernest Orlando, Lawrence Berkeley National Laboratory, Earth Sciences Di- Stoffa, P.L., Sen, M.K., 1991. Nonlinear multiparameter optimization using genetic algo-
vision, April1999. rithms: inversion of plane-wave seismograms. Geophysics 56, 1794–1810.
Rodriguez-Zuniga, J.L., Ortiz-Aleman, C., Padilla, G., Gaulon, R., 1997. Application of ge- Stork, C., Kusuma, T., 1992. Hybrid genetics autostatics: a new approach for large am-
netic algorithms to constrain shallow elastic parameters using “in situ” ground in- plitude statics with noisy data. Expanded abtrsacts 62nd SEG meeting, New
clination measurements. Soil Dynamics and Earthquake Engineering 16, 223–234. Orleans, USA, pp. 1127–1131.
Rucker, D.F., Ferre, P.A., 2005. Automated water content reconstruction of zero-offset Thurber, C.H., Ellsworth, W.L., 1980. Rapid solution of ray tracing problems in hetero-
borehole ground penetrating radar data using simulated annealing. Journal of Hy- geneous media. Bulletin Seismological Society America 70, 1137–1148.
drology 309, 1–16. Vassallo, M., Zollo, A., 2008. Depth and morphology of reflectors from the non-linear
Sambridge, M.S., 1990. Non-linear arrival time inversion: constraining velocity anom- inversion of arrival-time and waveform semblance data. Part I: method and appli-
alies by seeking smooth models in 3-D 1990 Geophysical Journal International cations to synthetic data. Geophysical Prospecting Vol. 56, 527–540. doi:10.1111/
102, 653–677. j.1365-2478.2008.00694.x.
Sambridge, M., 1999a. Geophysical inversion with a neighbourhood algorithm—I. Weber, Z., 2000. Seismic traveltime tomography: a simulated annealing approach.
Searching a parameter space. Geophysical Journal International Vol. 138, 479–494. Physics of the Earth and Planetary Interiors 119, 149–159.
Sambridge, M., 1999b. Geophysical inversion with a neighbourhood algorithm—II. Ap- Wu, Yih-Min, Zhao, Li, Chang, Chien-Hsin, Hsu, Ya-Ju, 2008. Focal-mechanism determi-
praising the ensemble. Geophysical Journal International Vol. 138, 727–746. nation in Taiwan by genetic algorithm. Bulletin of the Seismological Society of
Sambridge, M., Mosegaard, K., 2002. Monte Carlo methods in geophysical inverse prob- America 98 (2), 651–661. doi:10.1785/0120070115.
lems. Reviews of Geophysics 40 (3), 1–29.
Schwarzbach, C., Borner,, R.-Uwe, Spitzer, K., 2005. Two-dimensional inversion of di-
rect current resistivity data using a parallel, multi-objective genetic algorithm.
Geophysical Journal International Vol. 162, 685–695.

View publication stats