Building A 3D Initial Model For Full-waveform Inversion Via Genetic Algorithm And Deep Learning

Full-waveform inversion(FWI)is a powerful imaging tool generally implemented with gradient-based techniques.However,these techniques are very likely to get trapped into local minima,causing what is commonly known as cycle-skipping issues due to an inaccurate initial model.Unlike local optimization techniques,global optimization techniques such as genetic algorithm(GA)avoid all use of curvature information on the objective function;hence,they don’t require any derivative information and are easily customizable to work with any misfit function implementation.Recent advances in FWI have promoted the adoption of an optimal transport distance as a very convex misfit function between observed and modeled seismic data.However,the numerical computation of this distance requires the solution of a linear programming problem,which is prohibitively expensive when considering real field data applications,especially 3D seismic data.This dissertation proposes a strategy to tackle the cycle-skipping problem for estimating a 3D initial model based on GA while using a misfit function with entropy-regularized optimal transport distance followed by high-velocity body inversion using deep learning to improve the accuracy of the initial model.This strategy is illustrated through a series of numerical examples that are increasing in complexity,starting from misfit computation in schematic examples up to integrating GA,entropy-regularized optimal transport misfit function,and deep learning into more realistic full-waveform exercises,including the SEG/EAGE salt model.This strategy can obtain a 3D initial model that is accurate enough to match the traveltime of the diving waves within half of the cycle,being proved to be a robust solution for building a 3D initial model even with the presence of salt bodies.