Research On The Application Of Deep Learning Technology In Seismic Interpretation And Inversion

Seismic data interpretation and inversion are important techniques to obtain structural characteristics,sedimentary characteristics and reservoir parameters of underground media.In recent years,the deep learning technology in the field of artificial intelligence has made a lot of important progress in image recognition and classification,natural language processing and other aspects.Deep learning has been introduced to solve geophysical exploration probelms.This paper focused on the application of deep learning technology in seismic data interpretation and inversion,including the S-wave prediction of complex reservoirs,seismic horizon tracking and impedance inversion.S-wave velocity is one of the key parameters for prestack inversion.However,due to the high cost,most wells do not have the measured S-wave velocity.Empirical formula and petrophysical analysis methods have been used to obtain S-wave velocity,but these methods are always based on some assumptions,which impose some limitations on the accuracy and efficiency of S-wave velocity prediction for complex reservoirs.Taking complicated tight dolomite reservoir as a research subject,this paper proposes two methods to predict S-wave velocity from conventional logs by using deep fully connected neural network(FCNN)model and long and short time memory neural network(LSTM)model.FCNN model is easy to train and flexible,and LSTM model is suitable for processing serialization data and more stable.Both models have strong nonlinear expression ability and generalization.Finally,the trained model is used to predict S-wave velocity of other wells without the measurements.The two methods firstly select input features using crossplot and correlation analysis,and then select the well data with measured S-wave velocity to create training data.The FCNN and LSTM models for regression tasks were constructed respectively and trained.The prediction results on test well indicate that the two methods have achieved good performance for complex tight dolomite reservoirs.The accuracy and efficiency of S-wave velocity prediction are significantly improved compared with conventional rock physics methods and machine learning algorithms.The deep learning methods can provide accurate and reliable S-wave velocity of complex reservoirs for pre-stack inversion.Seismic horizons are essential data for structural analysis,inversion,seismic attribution analysis.Generally,horizon interpretation can be obtained by manual tracking and conventional automatic tracking techniques.Manual tracking is a time-consuming process.Conventional automatic tracking algorithms can improve the efficiency of horizon interpretation,but the accuracy of these algorithms is always be limited for target horizons with complex seismic reflections.In order to solve the problem,we proposed a CNN-based method to interpret seismic horizon.The method modified horizon-tracking to an image classification task to generate automaticly single or multiple horizons by using deep CNN model.The proposed method uses a fixed time window to scan over several seismic sections with the manual tracking horizon,and generate a set of image patches as the input of training data.According to the positional relations between the central point of each image patch and the target horizon,the image patch are assigned corresponding classifications as labels of training data.Then,A deep CNN model is constructed and trained to track target horizon in whole seismic volume.Results on test data indicate that the single and multiple horizons tracked by using this CNN-based method are in good agreement with the results of manual tracking,but the interpretation efficiency is significantly improved.Compared with the conventional automatic tracking techniques,the CNN-based method not only keep the tracking efficiency,but also significantly improves the accuracy of horizon interpretation under complex reflection characteristics.In addition,the research also proved that the CNN-based method is not strongly dependent on the complexity of input features.A CNN model trained by dataset with simple refelctions can still accurately interpret the target horizon under the condition of complex reflections.The application of deep learning technology in the field of seismic impedance inversion is still in the exploration stage.Most of the researches are based on supervised learning mode to employ deep neural networks for impedance inversion.The drawback of these methods is that a large number of measured impedances are required as training labels to obtain reliable inversion results,which seriously impose limitation on the application of deep learning technology in impedance inversion.In this paper,a method based on semi-supervised learning mode is proposed to estimate P-impedance inversion from post-stack data by using deep learning technology.The research designed an one-dimensional deep convolution neural network: INVNet model for P-impedance inversion.The prime advantage of this method is that training data does not need labels,which breaks through the limitation of deep learning on impedance inversion in the supervised learning mode.Post-stack seismic data and low-frequency data are the input of INVNet model.The output is defined as P-impedance.Synthetic seismic data are generated by using wavelet and the predicted P-impedance from the model output.Errors between the synthetic and real seismic data are calculated as waveform loss function.Errors between the predicted P-impedance and low-frequency data are calculated as low-frequency constraint loss function.The objective function of the INVNet model is constructed by assigning weight coefficients respectively for the two loss function terms,and is further optimized to train the model.In addition,an impedance hard constraint term is created and added the INVNet model to ensure the normal training and accelerate the convergence.The results on Marmousi forward model indicate that the proposed inversion method can estimate accurate P-impedance,and improve the efficiency of inversion.The weight coefficients of loss functions can affect the accuracy of inversion.The application in real seismic data shows that,compared with the conventional post-stack P-impedance inversion,the inversion results from the proposed method have higher resolution,better lateral continuity,and more consistent with the measured P-impedance of the test well.The research indicates that the proposed method can improves the accuracy and efficiency of P-impedance inversion...