Application Of Model-based Seismic Inversion Technology In The Prediction Of Thin Continental Sand Bodies

Deep learning technology is a hot technology in recent years,and deep learning-based formation inversion technology can improve the level of intelligence in interpretation.Because the interpretation of seismic strata is based on model inversion,the key to constructing the inversion model is the deep learning network.In order to solve the problems of low efficiency of traditional stratum interpretation and time-consuming and labor-intensive and other problems,the author proposes a Resnet reconstruction network structure,which adds 4 data feature pre-extraction layers before the resnet network structure,and trains the layer label to obtain Horizon inversion model,and predict the horizon based on the horizon inversion model,and on the basis of this prediction result,in order to solve the randomness of the model result and make it more suitable for the diversity of strata,perform cubic spline interpolation.Then solve the layer interpretation information that meets the geological target.Furthermore,a seismic inversion model is established in combination with logging data to obtain high-precision wave impedance.The specific implementation process is as follows:Aiming at the geological characteristics of the target area,this paper uses every three data as input data,and then predicts the horizon of the middle one,that is,the input is(400,3)(400 is the sampling point),and the output is(1,1);neural network Resnet18 is used as the main network layer for feature extraction,and then the self-built preprocessing layer and fully connected layer are added to form the total neural network;finally,cubic spline interpolation is used to correct the results.Since resnet18's original processing object is image data,the image data is composed of three channels of RGB,so the input data should also be composed of three passes.In the experiment,two methods are used to reconstruct the data volume: 1.The original data is deformed to obtain the three-channel data;2.The other seismic attributes are used for filling.The first method is relatively simple,has fast processing efficiency,does not require high calculation performance,and has poor results.The second type of data composition is complex,slow in efficiency,and the result is better.Resnet18 is used in the network because it is found in the experiment that as the number of network layers increases,the effect is slightly reduced.At the beginning,only the network structure of resnet18 was used,but the effect was found to be average.After adjusting the network structure,the network structure was gradually optimized,and finally the network reconstruction was completed.Finally,spline interpolation is performed to improve the prediction accuracy.The data will have an initial prediction result through the network,and then interpolate based on the result to achieve more accurate horizon prediction.