Deep Learning For 3-D Seismic Interpretation And Geological Modeling

Extracting and mining geological information from geophysical data is crucial for determining subsurface properties in oil and gas exploration and analyzing geodynamics in scientific problems.However,the increasing volume of data sets resulting from modern 3D exploration technologies has made traditional manual methods less effective in geological interpretation.To address these challenges,this paper explores multiple machine learning-based methods for structural interpretation,geological modeling,and reservoir prediction,which can extract geological information from various data sources.In fault surface construction,I propose two implicit surface methods based on least-squares optimization and machine-learning clustering.These methods consider the geometric relationships of all faults and compute an implicit scalar field by globally fitting the data.This improves the modeling capability for complex 3-D fault surfaces.The resulting comprehensive fault surface model can enhance subsequent geological structural modeling.In terms of structural interpretation,I propose two structural interpretation methods that are based on supervised learning and self-supervised learning.The first method focuses on predicting relative geological time volumes,while the second method involves self-supervised flattening of seismic volumes.To predict relative geological times from seismic amplitudes,I suggest adopting an end-to-end 3-D convolutional neural network.This trained network can automatically extract and analyze amplitude features from seismic data,accurately predicting relative geological volumes where each image pixel represents a seismic horizon.The simultaneous interpretation of horizons and faults in a global sense improves the geological correlation and spatial adaptability of the interpretation results.Seismic flattening aims to characterize structural deformations by mapping seismic images from the current space to a flattening domain where reflections are horizontally aligned.I propose a self-supervised flattening network that can accurately restore all horizons to an undeformed state and achieve consistent alignment of reflectors.The predicted deformations from the network can be used to calculate relative geological time volumes through linear transformation,enabling the interpretation of geological horizons.In structural modeling,I propose a supervised learning method to predict a structural model by using multi-source,heterogeneous,and sparse data.Rather than using numerical interpolation operators,this approach employs recursive convolutional kernels with adaptable parameters.This data-driven method can learn rich structural patterns from a large training dataset generated from simulation techniques,leading to improved representation of complex structural features.Furthermore,this method can incorporate various geological constraints into the modeling process when generating structural models from multi-source data,potentially surpassing the limitations of traditional modeling methods.With the structural models obtained from the aforementioned methods,I propose a data and model-driven reservoir inversion method by incorporating seismic and borehole data.This includes structure-guided borehole interpolation and deep learningbased reservoir prediction.These methods have wide applications in scientific research and oil and gas exploration,enabling accurate and efficient characterization of rock properties.