Multi-parameter Prestack Inversion Based On Deep Learning

Seismic inversion is the physical process of transforming seismic into petrophysical parameters using known seismic wave propagation theories and mathematical methods.Its mathematical essence is to establish a nonlinear relationship between seismic and petrophysical parameters.Deep learning can extract high-dimensional spatial feature information,establish a nonlinear mapping between input and output data,and strengthen the combination of deep learning and seismic inversion will help promote the automation and intelligent development of reservoir prediction.The use of deep learning for multi-parameter seismic inversion requires the construction of diverse dataset for reservoir parameter inversion.It completely relies on data-driven seismic inversion and lacks geophysical constraints,which makes the neural network easy to fall into over-fitting and the chemical ability is weak.In view of the need for effective and diverse training dataset for deep learning to carry out seismic inversion,this paper proposes a set of dataset construction procedures for reservoir parameter inversion.The process is based on the lithological properties and petrophysical parameters at the drilling location.Under the constraints of the low-frequency model,the Sequential Gaussian Co-Simulation is used to simulate the changes in the reservoir space,and the Elastic Distortion algorithm is used to modify the changes in the thickness of the formation space to make the petrophysical parameters The combination of vertical and horizontal directions is more abundant,and on this basis,the Xu-White model is used to convert petrophysical parameters into elastic parameters,and pre-stack seismic are synthesized according to the Zoeppritz equation to obtain reservoir-oriented multi-parameter inversion data with diverse characteristics set.For deep learning,multi-parameter seismic inversion is prone to over-fitting and weak generalization.Geophysical information is introduced into the deep neural network structure to improve the stability of pre-stack seismic inversion and reservoir parameters forecast accuracy.Based on the U-Net network,this paper considers the stability of the sparse distribution of reflection coefficients and the effectiveness of typical seismic attributes in reservoir parameters prediction during the network design process,and will have sparse reflection coefficients and seismic attributes sensitive to reservoir parameters.As the constraint information,the training process of inducing deep learning is similar to the physical process of seismic inversion,ensuring that the inversion result is more stable.The influence of geophysical information constraints,Batch Normalization layer,SNR of seismic data,and seismic wavelet frequency of inversion on the inversion results of elastic parameters and petrophysical parameters are quantitatively discussed,which provides for parameter optimization and popularization of deep learning inversion.The basis.The model test verifies that the P-wave velocity,S-wave velocity,density,porosity,and shale content obtained based on deep learning inversion have high accuracy,and the R~2 scores between the real model and the real model are 0.9373,0.9748,0.9758,0.8428,and 0.9003,respectively.The inversion test of the elastic and physical parameters of the actual work area shows that the inversion result matches well with the logging data and shows strong spatial continuity,which provides intelligent technical support for reservoir parameter prediction.