Coupled Multiscale Full Waveform Inversion
Parallel Systems, Full Waveform Inversion (FWI), Coupled Local Minimizers (CLM), Parallel Efficiency, Parallel Scalability.
Seismic reflection survey is the geophysical method better known and used. Oil and gas exploration is likely its greatest application. Its main objective is to generate an image of a region of the subsurface to identify structures of interest. In this type of survey, seismic waves are originated on surface. While transiting through the subsurface layers they are partially reflected. These reflections are then recorded on the surface. An important step in the processing of these seismic data is the velocity analysis. Its goal is to generate a model that inform wave propagation velocity in the region analyzed. In this context, Full Waveform Inversion (FWI) has become known for generating high precision models. Starting from an initial model, FWI simulates acquisition of seismic data. Iteratively, FWI seeks to minimize the difference between this calculated data and observed data. By using a local optimizer, often FWI does not converge to the global optimum model. Multiscale FWI seeks to reduce this problem with a smaller computational cost than the use of a global optimizer. Its strategy is invert the lower frequencies of the data initially and gradually add the higher frequencies. Thus, a model representing the major structures is obtained first and gradually the smaller structures are added. However, the model obtained from inversion of each frequency scale are used as initial approximation for the inversion of the next scale. This dependence reduces parallel escalability of this algorithm. Furthermore, it is common for real data to contain little low frequency information. Thus, inversion of lower frequencies may provide a poor approximation to the inversion of the next scales. Coupled Multiscale FWI (CMFWI) is proposed here. Its strategy is to play a Multiscale FWI in which the initial models used to invert the scales are independent of each other. This feature increases the efficiency of the parallel algorithm since it reduces the dependency between the inversions of the scales. To increase the likelihood of convergence to the global optimum, inversions scales are coupled, i.e., the inversion of each scale uses information about the other scales. This paper proposes to use the Coupled Local Minimizers (CLM) as CMFWI optimization method. However, CMFWI proposal is suitable for the use of other methods to minimize restrictions, like penalty methods.