Frequency-domain Full Waveform Inversion Based On Decreasing Random Shot Subsampling Method And Improved MLQN Method

With the development of underground reservoirs’ exploration, reservoirs which have simple geological background or convenient mining have been almost depleted. In order to satisfy the needs of social development for oil and gas resources, prospectors have to transfer their focus to the reservoir with complex structure setting. This is a challenge for accuracy of seismic exploration technology. In this paper, as a branch of seismic exploration technology, full waveform inversion method becomes the focus of research because it can accurately depict details of structure and lithology of targets with complex geological background. Some successful examples about marine seismic exploration prove that full waveform inversion provides quantitative basis for exploring oil-gas reservoirs which makes the success rate of drilling Wells reservoir increase dramatically.Full waveform inversion is a method which solves underground geophysical parameters by using all seismic waveform information. It uses the waveform information that include P wave, S wave, multiple wave, diffracted wave, etc. or amplitude, phase, traveling time. So its accuracy is higher. It is well known that the forward modeling is an important component part of inversion process. Forward modeling’s calculation accuracy and computation speed affect the quality and efficiency of inversion, so this paper compared some forward modeling methods. The finite difference method is simple and efficient, so this paper use it as a forward modeling method of full waveform inversion.While full waveform inversion takes its advantages, it still has some bottleneck problems. One of the problems is calculation amount. Especially for the 3D data volume, the calculation amount is generally too large to compute which results in no wide usage. Therefore, researchers put forward the encoded sources technology to reduce the calculation amount, but this method will introduce random crosstalk noise in the process of model updating. So, on the condition of guaranteeing the calculation accuracy, this paper firstly proposes decreasing random shot subsampling method which is used to efficiently calculate the full waveform inversion problem. On the condition of guaranteeing the calculation accuracy, this method not only reduces the calculation amount but also doesn’t introduce random crosstalk noise. It improves the efficiency of the full waveform inversion method.Full waveform inversion is a fitting process between simulated wave field value and the theory of wave field value. Its essence is an optimization problem. So, it is important to study optimization algorithm.The optimization algorithm which this paper studies based on the conventional memory quasi-newton algorithm(MLQN). Because the conjugate gradient method has super-linear convergence, this paper uses the Fletcher-Reeves(FR) conjugate gradient information to optimize the search direction of the conventional MLQN. Improved MLQN algorithm not only includes the gradient and model information but also includes conjugate gradient information. The new method improves precision of model obtained from the inversion. And, in the process of each update iteration of the model, this method does not increase the calculation amount. In this paper, the optimization method is used to single parameter and multi-parameter full waveform inversion. Inversion results is good.This paper mainly studies the observation system and optimization method. Through the analysis of the full waveform inversion process, it can be easily saw that the two research points are in the different two procedures. So, this paper combines the two research points in a full waveform inversion process and uses Overthrust model to test results. Study results show that it not only improves the computational efficiency of full waveform inversion but also improves the calculation precision of the inversion of the target body.