A new parallel simulated annealing algorithm for 1.5D acousticfull-waveform inversion

Xin Fu, Kristopher A. Innanen

Full-wave inversion (FWI) based on deterministic optimization (DO) methods is an appealing tool to detect the physical properties of the subsurface media, and increasing successful examples have been reported. However, the DO FWI is highly model-dependent, its success relies on a good starting model. To solve this problem, some researchers resort to the stochastic global optimization (SGO) methods that have shown the potential to alleviate the suffering of the model-dependent problem in FWI. Whereas, the SGO methods also have their own drawback that it needs to solve a great number of forward problems This is dramatically computationally expensive for the wave-equation-based FWI. In our study, we use a heuristic SGO algorithm, the very fast simulated annealing (VFSA) algorithm, to implement the constant-density acoustic FWI. To save computational time, we develop a new parallelization VFSA, in which the serial structure of VFSA is changed to some degree. Instead of updating the model parameter one by one in the same thread in a conventional serial VFSA, the parallel VFSA updates the N model parameter separately on N threads, in which the maximum efficient processor number is N. Since performing a 2D VFSA FWI directly without using any parameterization to reduce parameter number (dimension) is still prohibitive, we test the different FWIs (the DO FWI, the conventional serial VFSA FWI, and the new parallel VFSA FWI) on a 1.5D model, and for both VSP (vertical seismic profile) and surface seismic data. For each data, we use both an unbiased starting model crossing the true model and a biased starting model far away from the true model with the depth increase to investigate how the different FWIs rely on the starting model. The tests show that all FWIs for VSP seismic data are not too model-dependent, but the DO FWI for surface seismic data is mode-dependent and the VFSA FWIs can solve this problem well. Furthermore, to further save the computational cost, the data used for VFSA FWIs are multisource shot gathers. And all seismic data used are in time domain.