NLLoc performs earthquake locations in 3D models using non-linear search techniques.
Overview - Inversion Approach - EDT likelihood function - Oct-Tree Sampling - Grid-Search - Metropolis-Gibbs Sampling - Running the program-Input - Output - Processing and Display of results - [NonLinLoc Home]
The NLLoc program produces a misfit function, "optimal" hypocenters, an estimate of the posterior probability density function (PDF) for the spatial, x,y,z hypocenter location, and other results using either a systematic Grid-Search or a stochastic, Metropolis-Gibbs sampling approach.
The location algorithm used in NLLoc (Lomax, et al., 2000) follows the inversion approach of Tarantola and Valette (1982), and the earthquake location methods of Tarantola and Valette (1982), Moser, van Eck and Nolet (1992) and Wittlinger et al. (1993). The errors in the observations (phase time picks) and in the forward problem (travel-time calculation) are assumed to be Gaussian. This assumption allows the direct, analytic calculation of a maximum likelihood origin time given the observed arrival times and the calculated travel times between the observing stations and a point in xyz space. Thus the 4D problem of hypocenter location reduces to a 3D search over x,y,z space.
To make the location program efficient for complicated, 3D models, the travel-times between each station and all nodes of an x,y,z spatial grid are calculated once using a 3D version (Le Meur, 1994; Le Meur, Virieux and Podvin, 1997) of the Eikonal finite-difference scheme of Podvin and Lecomte (1991) and then stored on disk as travel-time grid files. This storage technique has been used by Wittlinger et al. (1993), and in related approaches by Nelson and Vidale (1990) and Shearer (1997). The forward calculation during location reduces to retrieving the travel-times from the grid files and forming the misfit function g(x) in, equation (3).
In addition, to save disk space and for faster calculation, a constant Vp/Vs ratio can be specified, and then only P travel-time grids are required for each station.
The Podvin and Lecomte (1991) algorithm and related methods use a finite-differences approximation of Huygen's principle to find the first arriving, infinite frequency travel times at all nodes of the grid. The algorithm of Podvin and Lecomte (1991) gives stable recovery of diffracted waves near surfaces of strong velocity contrast and thus it accurately produces travel times for diffracted and head waves. A limitation of the current 3D version of the method is a restriction to cubic grids. This may lead to excessively large travel-time grids if a relatively fine cell spacing is required along one dimension since the same spacing must be used for the other dimensions. This can be a problem for regional studies where a fine node spacing in depth is necessary, but the horizontal extent of the study volume can be much greater than the depth extent. Thus a modification of the travel times calculation to allow use of an irregular grid would be very useful.
After the travel times are calculated throughout the grid, the NonLinLoc program uses the gradients of travel-time at the node to estimate the take-off angles at each node. Two gradients are estimated for each axis direction x, y, and z - one Glow between the node and its preceding neighbour along the axis, and a second Ghigh between the following neighbour and the node. The total gradient Gaxis along an axis is the mean of these two gradients; the total gradient along the three axes determines the vector gradient of travel-time. The direction opposite to the vector gradient of travel-time gives the ray take-off angles for dip and azimuth. An estimate of the quality of the angle determination is given by a comparison of the magnitudes and signs of Glow and Ghigh. If these two values are not similar, then there may be two rays which arrive nearly simultaneously at the station, and the take-off angle determination at the node may be unstable.
The x,y,z volume used for grid-search or Metropolis-Gibbs location must be fully contained within the 3D travel-time grids. This limits the largest station distance that can be used for location since the 3D travel-time computation and the size of the output time-grid files grow rapidly with grid dimension. However, for location in flat-layered media, the travel times can be stored on very compact 2D grids, and readings for "regional" stations far from the search volume can be used.
Except for TRANS GLOBAL mode, the NLLoc program uses a "flat earth", rectangular, left-handed, x,y,z coordinate system (positive X = East, positive Y = North, positive Z = down). Distance units are kilometers, and many input/output distance quantities can be expressed in rectangular or geographic (latitude and longitude) coordinates.
In TRANS GLOBAL mode, the NLLoc program uses a "spherical earth", longitude,latitude,depth
coordinate system (positive X = East, positive Y = North, positive Z =
down). Longitude and latitude units are degrees, depth is in
kilometers, and most input/output distance quantities are expressed in
geographic (latitude and longitude) coordinates.
See the book chapter Probabilistic earthquake location in 3D and layered models: Introduction of a Metropolis-Gibbs method and comparison with linear locations (Lomax, et al., 2000) for further information on the NonLinLoc location algorithms.
The earthquake location algorithm implemented in the program NLLoc (Lomax, et al., 2000) follows the probabilistic formulation of inversion presented in Tarantola and Valette (1982) and Tarantola (1987). This formulation relies on the use of normalised and unnormalised probability density functions to express our knowledge about the values of parameters. Thus, given the normalised density function f(x) for value of a parameter x, the probability that x has a value between X and X+DX is
In geophysical inversion we wish to constrain the values of a vector of unknown parameters p, given a vector of observed data d and a theoretical relationship q (d,p) relating d and p. When the density functions giving the prior information on the model parameters rp(p) and on the observations rd(d) are independent, and the theoretical relationship can be expressed as a conditional density function q (d|p)mp(p), a complete, probabilistic solution can be expressed as a posterior density function (PDF) sp(p) (Tarantola and Valette, 1982; Tarantola, 1987)
where mp(p) and md(d) are null information density functions specifying the state of total ignorance.
For the case of earthquake location, the unknown parameters are the hypocentral coordinates x = (x, y, z) and the origin time T, the observed data is a set of arrival times t, and the theoretical relation gives predicted travel times h. Tarantola and Valette (1982) show that, if the theoretical relationship and the observed arrival times are assumed to have gaussian uncertainties with covariance matrices CT and Ct, respectively, and if the prior information on T is taken as uniform, then it is possible to evaluate analytically the integral over d in (2) and an integral over origin time T to obtain the marginal PDF for the spatial location, s(x). This marginal PDF reduces to (Tarantola and Valette, 1982; Moser, van Eck and Nolet, 1992)
In this expression K is a normalisation factor, r(x) is a density function of prior information on the model parameters, and g(x) is an L2 misfit function. is the vector of observed arrival times t minus their weighted mean, is the vector of theoretical travel times h minus their weighted mean, where the weights wi are given by
Furthermore, Moser, van Eck and Nolet, 1992 show that the maximum likelihood origin time corresponding to a hypocentre at (x, y, z) is given by
The posterior density function (PDF) s(x) given by equation (3) represents a complete, probabilistic solution to the location problem, including information on uncertainty and resolution. This solution does not require a linearised theory, and the resulting PDF may be irregular and multi-modal because the forward calculation involves a non-linear relationship between hypocentre location and travel-times.
This solution includes location uncertainties due to the spatial relation between the network and the event, measurement uncertainty in the observed arrival times, and errors in the calculation of theoretical travel times. However, realistic estimates of uncertainties in the observed and theoretical times must be available and specified in a gaussian form through Ct and CT, respectively. Absolute location errors due to incorrect velocity structure could be included through CT if the resulting travel time errors can be estimated and described with a gaussian structure. Estimating these travel time errors is difficult and often not attempted. When the model used for location is a poor approximation to the "true" structure (as is often the case with layered model approximations), the absolute location uncertainties can be very large.
The NLLoc grid-search algotithm systematically determines the posterior probability density function s(x) or the "misfit" function g(x) over a 3D, x,y,z spacial grid. The NLLoc Metropolis-Gibbs sampling algorithm attempts to obtain a set of samples distributied according to the posterior probability density function s(x).
The grid-search s(x) grid, samples drawn from this function, or the samples obtained by the Metropolis-Gibbs sampling, form the full, non-linear spatial solution to the earthquake location problem. This solution indicates the uncertainty in the spatial location due to picking errors, a simple estimate of travel-time calculation errors, the geometry of the observing stations and the incompatibility of the picks. The location uncertainty will in general be non-ellipsoidal (non-Gaussian) because the forward calculation involves a non-linear relationship between hypocenter location and travel-times.
Because it is difficult or impossible to obtain, a more complete estimate of the travel-time errors (or, equivalently, a robust estimate of the errors in the velocity model) is not used. This is a serious limitation of this and most location algorithms, particularly for the study of absolute event locations.
The PDF may be output to a 3D Grid and a binary Scatter file (see Output below). PDF values are also used for the determination of weighted average phase residuals (output to a Phase Statistics file), and for calculating location confidence contour levels (see Output below), and "Traditional" Gaussian estimators (see below).
The maximum likelihood (or minimum misfit) point of the complete, non-linear location PDF is selected as an "optimal" hypocentre. The significance and uncertainty of this maximum likelihood hypocentre cannot be assessed independently of the complete solution PDF. The maximum likelihood hypocenter parameters are output to the NNLoc, ASCII Hypocenter-Phase File (HYPOCENTER, GEOGRAPHIC and QUALITY lines), and to the quasi-HYPOELLIPSE format and HYPO71 format files. The maximum likelihood hypocenter is also used for the determination of ray take-off angles (output to a HypoInverse Archive file), for the determination of average phase residuals (output to a Phase Statistics file), and for magnitude calculation. The ray take-off angles can be used for a first-motion fault plane determination.
"Traditional" gaussian or normal estimators, such as
expectation E(x) and covariance matrix C may be
from the gridded values of the normalised location PDF or from samples
of this function (e.g. Tarantola and
1982; Sen and Stoffa,1995). For the
case with nodes at xi,j,k,
volume of a grid cell. For N samples drawn from the PDF with locations
where the PDF values s(xn)
are not required since the samples are assumed distributed according to
the PDF. For both cases, the covariance matrix is then given by
The Gaussian estimators are output to the NNLoc, ASCII Hypocenter-Phase File (STATISTICS line).
The 68% confidence ellipsoid can be obtained from
value decomposition (SVD) of the covariance matrix C, following
Press et al. (1992; their sec. 15.6 and eqs. 2.6.1 and
The SVD gives:
where U = V are square, symmetric matrices and wi are singular values. The columns Vi of V give the principle axes of the confidence ellipsoid. The corresponding semi-diameters for a 68% confidence ellipsoid are Ö (3.53wi), where 3.53 is the Dc2 value for 68.3% confidence and 3 degrees of freedom.
The gaussian estimators and resulting confidence ellipsoid will be good indicators of the uncertainties in the location only in the case where the complete, non-linear PDF has a single maximum and has an ellipsoidal form.
EDT description page
Oct-Tree description page
The grid-search algorithm performs successively finer, systematic grid-searches within a spatial, x,y,z volume to obtain a misfit function, an optimal hypocenter and an estimate of the posterior probability density function (PDF) for hypocenter location.
The Grid-Search location is based on a nested grid search using one or more location grids as specified by LOCGRID statements in the Input Control File. The first LOCGRID statement specifies a specific initial search grid with fixed size, number of nodes and location. Subsequent LOCGRID statements specify the size and number of nodes for subsequent, nested grids; the location of these nested grids is usually set automatically in one or more of the x,y,z directions.
For each location grid, the location quality (misfit or PDF value) at every node is obtained. For each node, the travel-times for each observation are obtained from the corresponding travel-time grid file and the PDF s(x), or misfit value g(x) is calculated using the equations given above in the Inversion Approach section. These location quality values are saved to a 3D grid file if requested. If there is a subsequent nested grid, its position (for the directions with automatic positioning) is set so that it is centered on the maximum PDF node (or, equivalently, the minimum misfit node) of the current grid.
The initial location grid must be fully contained within the travel-time grid files corresponding to a given observation for that observatoin to be used in the location. Subsequent location grids, even if their position is set automatically, must be fully contained within the initial grid. The NLLoc program will attempt to translate a nested grid that intersects a boundary of the initial grid so that it is contained inside of the initial grid; if this is not possible the location will be terminated prematurely.
For every node of each location grid, the grid-search algorithm must obtain travel-times for every observation. These times are stored on disk in 3D travel-time grid files which may be very large. It would be extremely time consuming to read these times one by one directly from the disk files, but there is also not enough space in general to fully read all the relevant 3D grid files into memory. However, the grid search is performed systematically throughout each location grid with the x index varying last. Thus, it is adequate to have 2D planes or "sheets" corressponding to the current x index available in memory at any one time. This approach is used by the grid-search algorithm. Sheets of data with a given x index are read from the 3D travel-time grid files as large blocks of bytes, which is very fast in comparison to reading the same number of data values individually.
The Metropolis-Gibbs algorithm performs a directed random walk within a spatial, x,y,z volume to obtain a set of samples that follow the 3D PDF for the earthquake location. The samples give and estimate of the optimal hypocenter and an image of the posterior probability density function (PDF) for hypocenter location.
The Metropolis-Gibbs search proceedure to obtain samples of a PDF is based on the algorithm of Metropolis et al. (1953) for the simulation of the distribution of a set of atoms at a given temperature. The Metropolis-Gibbs algorithm used here is similar to the "Metropolis" algorithm described in Mosegaard and Tarantola (1995) and the "Gibbs sampler" with temperature T=1 described in Sen and Stoffa (1995; sec 7.2). It may be considered as a version of Metropolis simulated annealing Kirkpatrick et al. (1983) where the temperature parameter is a constant determined by the covariance matrix for the observational and forward problem uncertainties. Thus the algorithm does not "anneal" or converge to an optimal solution, but instead produces a set of samples which follow the posterior PDF for the inverse problem.
The Metropolis-Gibbs sampler used in the program NonLinLoc for earthquake location consists of a directed walk in the solution space (x, y, z) which tends towards regions of high likelihood for the location PDF, s(x) given by equation (3). At each step, the current walk location xcurr is perturbed by a vector dx of arbitrary direction and given length l to give a new location xnew. The likelihood s(xnew) is calculated for the new location and compared to the likelihood s(xcurr) at the current location. If s(xnew) ³s(xcurr), then the new location is accepted. If s(xnew) < s(xcurr), then the new location is accepted with probability P = s(xnew) / s(xcurr). When a new location is accepted it becomes the current location and may be saved as a sample of the location PDF.
In earthquake location, the dimensions of the significant regions of the location PDF can vary enormously and are not known a priori. It is important to choose an initial step size large enough to allow global exploration of the search volume, and to obtain a final step size that gives good coverage of the location PDF while resolving details and irregular structure of the PDF. The NonLinLoc Metropolis-Gibbs sampler uses three distinct sampling stages to determine adaptively an optimal step size l for the walk:
It is important to set the parameters for the directed walk so that (1) during the learning and equilibration stages the walk approaches and reaches the high likelihood regions of the location PDF, and so that (2) by the saving stage a suitable, relatively small, fixed step size has been obtained to accurately explore and image the PDF.
The NonLinLoc Metropolis-Gibbs sampling algorithm is initialised as follows:
The rejection by the algorithm of new walk locations for a large number of consecutive tries (the order of 1000 tries) may indicate that the last "accepted" sample falls on a sharp likelihood maxima that is narrower than the current step size. To allow the search to continue in this case, the new location is accepted unconditionally and the step size is reduced by a factor of two.
In the case that the size of the location PDF is very small relative to the search region, the algorithm may fail to locate the region of high likelihood or obtain an optimal step size. In this case the size of the search region must be reduced or the size of the initial step size adjusted. A more robust solution to this problem may be to add a temperature parameter to the likelihood function, as with simulated annealing. This variable parameter could be set to increase the effective size of the PDF during the learning and equilibration stages so that the region of high likelihood is located efficiently, and then set to 1 during the saving stage so that the true PDF is imaged.
Synopsis: NLLoc InputControlFile
The NLLoc program takes a single argument InputControlFile which specifies the complete path and filename for an Input Control File with certain required and optional statements specifying program parameters and input/output file names and locations. See the NLLoc Statements section of the Input Control File for more details. Note that to run NLLoc the Generic Statements section of the Input Control File must contain the CONTROL and TRANS (Geographic Transformation) statements.
In addition, the NLLoc program requires:
The names, locations and other information for these files is specified in the NLLoc Statements section of the Input Control File.
The location results can be output for single event and summary (all events) as:
* these output types are only generated for grids where
PDF is calculated.
! these output types are only written to single event files
The location results can also be output as summary (all events) files containing:
Single event and summary files are only saved for specific nested search-grids as specified in the LOCGRID statement in the Input Control File.
The location results for one or more events can be combined with the program LocSum to produce output such as a comprehensive, summary Hypocenter-Phase File, a binary Scatter File, and a set of simple ASCII format Scatter samples files.
The a comprehensive, summary Hypocenter-Phase File forms the input for the Java applet SeismicityViewer for interactive, 3D display of event locations on an internet browser or appletviewer.
The location results for a single event or the output files produced by the program LocSum can be post-processed with the program Grid2GMT to produce a GMT command script for plotting misfit, PDF and location "cloud" results using the GMT plotting package.
Back to the NonLinLoc site Home page.