A Study Of Neural Networks-based Initial Velocity Model Building Method For Full Waveform Inversion

In the field of geophysical exploration,the velocity of seismic wave propagation in the geology is the key parameter to be acquired.Obtaining an accurate initial velocity model is a key prerequisite for full waveform inversion imaging.Compared with other imaging methods that use only partial wave field information,the full waveform inversion algorithm uses full wave field parameter information and can accurately recover the geological structure even under complex geological conditions.However,the full waveform inversion algorithm is not well used in practical industry,mainly because it relies heavily on the initial velocity model,and if the initial model is not accurate enough,it is easy to fall into periodic oscillations and get multiple extremes.In recent years,with the hot rise of deep learning,it has achieved attractive results in many fields.In the field of geological exploration,deep learning can also be used to solve problems that were previously difficult to solve by traditional algorithms.To address the problem that the full waveform inversion algorithm relies heavily on the initial velocity model,this paper proposes a scheme to build a velocity model directly in seismic data using convolutional networks,and also establishes a web application platform that can be used for the corresponding work such as geological data interpretation.The main work of this paper is as follows.(1)In this thesis,velocity models are extracted from geological data and a large number of snapshots of seismic wave fields are generated as training datasets by simulating orthogonal fluctuations,while label sets are made with source velocity models as mapping labels.After preparing the dataset,the network structure of Attention-R2U-NET is designed to learn features and build velocity models directly from seismic data by combining the features of recurrent network and recursive network based on U-Net network,replacing the original nodes,and adding attention mechanism at the same time.(2)In this thesis,a GAN-FWI network based on adversarial neural network is proposed,and the initial velocity model is parameterized by replacing the generator and discriminator in the generative adversarial network,using the weight data to automatically capture the significant features in the initial model and use them as a priori information.After that,the discriminator full waveform inversion algorithm is used to update the velocity model,and since the value of this velocity model has a mapping relationship with the value of the network weights,the network weights can be updated by the velocity model,which enables the generator network to generate realistic velocity models.(3)In this thesis,a deep learning-based Web platform for seismic phase rock formation identification is designed.The platform provides an automated recognition service for geological data,and the trained network model is deployed on the server,which can recognize the input and return the results,making the algorithm processing transparent.