Machine Learning Assisted Seismic Interpretation And Inversion

The explosion of data types and volumes we are experiencing in oil and gas field is stretching the capabilities of traditional geophysical workflows.The rising of artificial intelligence promoted machine learning applications in each step of geophysical exploration,which bring us new opportunity and challenges.Based on machine learning approaches,this dissertation discussed classification and regression problems in geophysics,including first-arrival picking,log-facies classification,and seismic inversion.We proposed an automated first arrival picking method by image segmentation.Considering the features of seismic data,we first improved the U-Net architecture to preserve good performance and reduce computation cost simultaneously.Then we introduced spatial information into model training,which can help to eliminate picking outliers and achieve stable performance.Field examples demonstrate its superiority over traditional STA/LTA and other similar machine learning methods.U-Net also shows its potential to correct human picking error and reduce human labor.Furthermore,U-Net can provide accurate picking even for seismic data with ultra-low signal-to-noise ratio in a fast and stable way,which is crucial for its field application.For log facies classification,we first applied different supervised learning methods and field examples show that the superiority of XGBoost over support vector machine and random forest.Then,we proposed a semi-supervised learning method for log facies classification with small labeled datasets.Data augmentation is introduced to avoid overfitting and a novel loss function that contains unlabeled data loss to improve the model's ability to generalize.Field examples indicate that the improved semi-supervised learning method can outperform the supervised methods.It can also match well with manual interpreted facies and effectively reduce the human labor.We proposed a novel inversion method that can automated update and expand training datasets,which can help to solve the shortage problem of labeled samples during seismic inversion.Our workflow first chooses neighbor traces around each well as the initial training dataset,then train and update the model from near to far trace by trace.A data augmentation strategy is also applied to avoid overfitting,which is derived from convolution model.To reasonably arrange the influence weights of different wells to each seismic trace,both azimuth and distance are used to compute the well weights.The proposed inversion method shows solid results on both synthetic and field examples for predicting acoustic impedance.It can help to characterize channels and spatial information of thin-layer reservoir.Accurate matching between well-log interpretation and impedance also demonstrate its high potential in reservoir prediction.