For a particula http www apexbt

For a particular workflow of FWI or sensitivity/resolution analysis, a user must set a parameter file that specifies all general information that will not change throughout iterations of full waveform pmsf (if there are any) and from which locations of all files and directories used by the workflow can be inferred. Therefore, it is called the main parameter file (Fig. 2, ) and it is required as input to almost all ASKI executables. Along with some conventions on nomenclature, all files required by an executable can thus be located on the file system. All seismic sources and receivers involved in this workflow need to be specified by text files in a simple pre-defined format. Since ASKI works in the frequency domain, any waveform data to be inverted or filters used in the operations need to be Fourier-transformed at specifically chosen discrete frequencies. ASKI provides executables for these tasks (Fig. 2, ) and supports basic formats for seismic data such as text trace files and Seismic Unix [24]. The user needs to choose a forward code by which the wavefields required for kernel computation should be computed. The particular choice may depend on the geophysical complexity that should be accounted for in forward modeling (e.g. 1D or 3D acoustic or visco/poro-elastic medium, local Cartesian model domain or accounting for Earth’s curvature/gravity/rotation). Sufficient experience is necessary how to operate the forward code in general, as well as its specific features of producing output for ASKI. These output files must be written to a designated directory referred to by the main parameter file (Fig. 2, ). This requires large amounts of storage and significant output operations, but combining the wavefields for different source–receiver pairs when computing kernels by the SI method may result in an overall optimized number of simulations to be done, dependent on the involved number of sources and receivers [9]. ASKI currently supports computation of spectral waveform kernels for isotropically elastic model parametrizations. Kernels can be computed selectively only for those source–receiver combinations for which there are data in the dataset, possibly at individual frequencies. Calling the respective excecutable, the wavefield files are read in and the computed kernel files are stored to their designated directory (Fig. 2, ). By the formulae for isotropically elastic waveform kernels [11, Eq. A5], eighteen displacement and strain components are superposed to just three kernel values (for two elastic constants and density). Additionally the kernels are pre-integrated onto the coarser inversion grid. That is why they require significantly less disk space than the wavefields on which they are based (only around 3%, for instance, in the example inversion shown by [11, table 1]). Once kernels are computed, they can be used for deriving an update of the isotropically elastic earth model parameters in an iteration of FWI (Fig. 2, –). Smoothing and/or damping conditions can be added as additional equations to the linear system of sensitivity equations. These equations relate the unknowns (i.e. the model update values) in a specific way, e.g. forcing them to be small (damping) or requesting them to represent some average of their neighboring values (smoothing). For every model update value, the particular constraints are in general allowed to vary with location inside the model domain. Therefore, the intensities of smoothing and damping can be reduced in areas of good data coverage, thus increasing the influence of the seismic data in places where they have more resolutional power. Applying this kind of regularization is independent of the preceding stages of solving the forward problem and computing kernels, which are also much more computationally expensive than deriving a model update by solving an overdetermined system of equations in a least-squares sense. Thus, a user can play with smoothing and damping intensities and derive different model updates at relatively low costs. In addition, particular data can be down-weighted or discarded at this stage in order to reduce unwanted influence by data that can only be fitted insufficiently or to counteract model artifacts evoked e.g. by earthquake clusters. This provides the user with additional flexibility for deriving a satisfactory updated model.