Seismic Inversion Of Pre-stack Elastic Parameters Based On Deep Learning

Seismic inversion is an important method for quantitatively interpreting the physical properties of subsurface rocks.Among them,prestack inversion,due to its ability to preserve seismic reflection amplitudes that change with offset or incidence angle,can provide more geophysical parameters reflecting the lateral variation of reservoirs,and is of great significance for the fine characterization of reservoir spatial distribution and physical properties in complex oil and gas fields.However,traditional prestack inversion methods suffer from issues such as non-uniqueness and uncertainty due to the limited bandwidth of seismic data,inaccuracies in forward physical models,and unknown seismic wavelets.In recent years,with the development of deep learning and big data,more and more deep learning methods have been applied in the field of geophysical exploration and have shown remarkable potential.Based on this,this thesis attempts to apply deep learning methods to prestack elastic parameter inversion to reduce the impact of issues such as non-uniqueness and the difficulty in expressing complex geological features in traditional methods on the inversion results.This thesis proposes an end-to-end multi-task deep learning model that integrates fully convolutional neural networks(FCN)and bidirectional gated recurrent units(BiGRU)to achieve intelligent inversion from "seismic data to elastic parameters".Using the same variance uncertainty to define the loss function of the multi-task network model avoids manually adjusting the weight coefficients of different tasks.Secondly,based on the FCN-Bi GRU network model,double supervised elastic parameter inversion method based on attention mechanism is proposed based on the classical autoencoder architecture.In this method,the decoder establishes a mapping relationship between logging data and seismic data,and adds seismic data loss matching terms to the model’s loss function,using seismic waveform matching constraints to reduce the inversion’s solution space and further reduce the non-uniqueness of data-driven inversion.At the same time,the attention mechanism is incorporated into the encoding network to improve the prediction accuracy of the inversion results.Using the FCN-Bi GRU network model,tests on the Marmousi model showed that this method can maintain the rock physics relationship between different elastic parameters during the inversion process and has accurate and fast inversion capabilities.The influence of factors such as noise level,well position,and seismic data frequency on this method is also explored and understood to some extent.Using the double supervised method,this method also has good inversion capabilities even when seismic data is disturbed by noise,and the inversion results better conform to the law of seismic wave propagation.Finally,attempts to apply the above method to actual data show its practicality and reliability.