S-Wave Velocity Prediction,Seismic First Break Picking,and TOC Data Augmentation Based On Deep Neural Networks

Deep neural networks(DNNs)are artificial neural networks that build complex patterns and stronger non-linear relationships between the input and the output through multiple hidden layers,which makes them suitable for various geophysical applications.DNNs are a powerful tool that can provide new insights and solutions for various geophysical problems,such as geophysical data prediction,classification,and augmentation.DNNs are robust generalization algorithms in the presence of new data and can be efficiently applied to large volumes of seismic data for image segmentation.DNNs exceed in automatic feature extraction,learn complex relationships,and capture nonlinear and temporal dependencies for properties prediction.Furthermore,DNNs can be used for data augmentation to improve the robustness of the model and reduce overfitting and the need for new data,especially when available data are limited or difficult to obtain due to time or economic factors.In this dissertation,DNNs are implemented to address the limitations of existing methods for shear wave(S-wave)velocity estimation,seismic first break(FB)picking,and total organic carbon(TOC)content data augmentation.S-wave velocity is an important elastic parameter for seismic inversion,reservoir characterization,and geotechnical analysis.S-wave velocity can help to identify matrix minerals,formation porosity,and the type of fluid contained in a geological formation,especially in gas reservoirs.However,the measured S-wave velocity is often unavailable in most wells due to technical or economic reasons.Therefore,numerous conventional methods,such as empirical formulas and rock physics models,have been proposed to estimate the S-wave velocity and understand the non-linear relationships between the elastic rock properties.Though,empirical formulas are valid locally and require timeconsuming calibration for global application.At the same time,rock physics models require many difficult-to-obtain parameters to achieve reliable results(e.g.,mineralogy content,pore aspect ratio,and formation pressure).Conversely,recurrent neural networks(RNNs)can automatically and efficiently map the unseen non-linear relationship between well-log data and S-wave velocity,especially in complex geological areas.Though RNNs are promising,they have limitations in predicting long-term sequences,especially when the spatiotemporal variations of S-wave velocity are not considered.FB picking is a critical step in seismic data processing that provides information on the velocity of the seismic waves,infers the properties of the geological structures on the near surface,and improves the seismic imaging of the subsurface structures.However,FB picking by hand is a labor-intensive,time-consuming,and low-efficient task,particularly for large seismic datasets.At the same time,conventional methods for FB picking are prone to errors,particularly in noisy or complex geologic settings,such as seismic faults,lithological unconformities,or abrupt lateral velocity variations.Conversely,U-Net can exceed in automation,accuracy,and efficiency for FB picking.U-Net can process large volumes of seismic data much faster than manual and conventional methods.U-Net learns complex patterns in seismic data and identifies subtle features that may be difficult for human analysts to detect.However,U-Net is also sensitive to noise and has difficulties managing complex geological structures with high heterogeneity.TOC content is important in hydrocarbon exploration because it helps to evaluate the thermal maturity,quality,and source of the organic matter for hydrocarbon generation on sedimentary rocks.Thus,higher TOC content values indicate greater potential for hydrocarbon formation.However,continuous and reliable TOC content data are challenging due to costly core sampling and specialized analytical techniques.Therefore,predicting TOC content with few samples is challenging(i.e.,few-shot learning),especially when lacking high TOC content data.Generative adversarial networks(GANs)are one alternative for TOC content data augmentation due to their flexibility in generating reliable results that resemble the measured TOC content data distribution.However,GANs are difficult to train,especially when TOC content data are scarce or not presentative of the overall dataset and have limitations to directly controlling TOC content data generation.Therefore,in order to address the current limitations of DNNs,the consideration of spatiotemporal dependencies for S-wave velocity prediction,the accuracy of seismic FB picking in noisy scenarios,and the lack of control over the augmentation of TOC content data with few high TOC content samples,three approaches are introduced.First,a method for predicting S-wave velocity based on graph convolutional networks(GCNs)and recurrent neural network(RNN)variants are proposed.GCN extracts intrinsic spatial and temporal relationships,identifies patterns,and determines relevant information contributing to the prediction.While long-short-term memory(LSTM)and bidirectional gated recurrent units(Bi GRUs)forecast long-term data sequences considering temporal-depth relationships.The well-log curves are used to construct the graph topology,the well-log data are the node features,and their correlation coefficients are the weighted edges.The workflow implements the total information coefficient(TIC)as an external knowledge to map non-linear relationships among the well-log data to capture robust invariant correlations with depth,increases the prediction accuracy,and improves the interpretability of the model.Moreover,an unsupervised graph neural network(GNN)is applied to identify and remove outliers before training based on the proximity of nearby samples through message passing.Furthermore,the complete ensemble empirical mode decomposition with additive noise(CEEMDAN)is implemented to construct time-frequency features sensitive to specific stratigraphic and geological characteristics,thus increasing the accuracy of the model.Finally,results show that the proposed approach predicts the S-wave velocity with high accuracy and reliability compared with Castagna’s equation and alternative DNNs.Second,a method for FB picking based on U-Net variants is proposed.U-Net divides the seismic shot gather into two regions,before and after the first arrival,and extracts the FB from the boundary.Then,the performance of FB picking is evaluated using U-Nets with different architectures and complexities,including U-Net,wide U-Net,U-Net++,and attention U-Net.Additionally,an interquartile range(IQR)seismic signal enhancement and apparent velocity constraint(AVC)are implemented as pre-and postprocessing techniques to increase the accuracy of the extracted FB.The seismic signal enhancement combines geometrical spreading correction,IQR clipping,and RMS normalization to increase the seismic energy and balance the signal-to-noise ratio(SNR),thus improving the classification performance of U-Net.AVC improves the segmentation accuracy and reduces the misclassification of pixels in the segmented shot gather by establishing upper and lower boundaries around the first arrival.The classification of pixels within the constraints is preserved,while the outside is rectified.Results from land seismic datasets show that the proposed workflow can accurately extract the seismic FB from 2D shot gathers,for both training and blind datasets,in the presence of coherent and random noise and regions with low seismic energy in the middle and far offsets.Third,conditional generative adversarial networks(CGANs)are introduced to generate reliable TOC content data with few high TOC content samples.CGAN uses petroleum geochemical information as the input conditions to guide the generation of new samples.CGAN trains a generator network to extract the distribution of the measured TOC content samples and a discriminator network to evaluate the quality of the generated TOC content samples.Furthermore,the Mahalanobis distance constraint eliminates outliers and balances the distribution based on their distance from the center of the distribution.Results show that the proposed workflow generates reliable TOC content samples that preserve the characteristics of the measured TOC content distribution and their intrinsic elastic rock relationships,thus increasing their correlation coefficient.