Full Waveform Inversion Based On Optimal Transport And Recurrent Neural Networks

Seismic exploration is an important means to find mineral resources,like oil and gas.Effective exploration is related to the development of the national economy and the security of national energy.With the difficulty of exploration increasing,it is particularly critical to improve the quality of imaging for the detailed study of underground stratigraphic structure.At present,the full waveform inversion is one of the most accurate seismic exploration methods,which uses predicted seismic data and observed seismic data to establish the objective function with some metric.Then it updates the predicted model by minimizing the objective function through the optimization algorithm.Although the full waveform inversion has the potential to reveal the stratigraphic structure under the complex geological exploration background with high-resolution,it still has some trouble to work well.For instance,cycle-skipping issues lead to the strong dependence of the inversion on the initial model and the large computation of the inversion lead to the low efficiency of inversion,which affects the accuracy and convergence speed of the full waveform inversion.In fact,it is difficult to acquire the initial model with high accuracy,while different objective functions have different requirements for the initial model.In addition,optimization algorithms have an impact on the inversion efficiency and accuracy.Therefore,this dissertation focuses on the selection of the objective functions and optimization algorithms with the aim of advoiding suffering from cycle-skipping and improving the efficiency of the full waveform inversion.This dissertation uses a recurrent neural network to carry on the acoustic wave equation forward modeling in time domain,and according to the theory of optimal transport obtains different objective functions solved by the optimization algorithms of deep learning.So the dependence of inversion algorithm on the initial model is improved and the efficiency of inversion becomes better.Firstly,in the framework of Tensor Flow,full waveform inversion combined with optimal transport and recurrent neural network is feasible.Secondly,the influence of different metric methods and optimization algorithms on inversion is explored through numerical experiments,where the objective function based on quadratic Wasserstein distance with a linear positive transformation using adam method is a good candidate for inversion.It not only improves the cycle-skipping phenomenon,reduces the dependence on the initial model,but also makes the inversion more efficient.At the same time,considering the impact of different batch sizes on the above inversion method,and compared with the conventional(full batch)full waveform inversion,it proves that the mini-batch(random)inversion strategy can not only ensure the inversion accuracy,but also need lower requirements for the computing resources,which is a well-performing inversion method.Finally,add noise to observed data.Although the quadratic Wasserstein metric based on linear positive transformation is supposed to recover the general velocity structure,the objective function based on the least-square metric adopts the cascade inversion method has a better result.