LAUSR

Abstract We propose a machine learning-based seismic spectral attribute (SSA) analysis to delineate the thickness of a tight-sand reservoir in the Sulige gas field of central Ordos Basin, western China. In our workflow, we first implement the seismic spectral decomposition by using the continuous wavelet transform (CWT) with the generalized Morse wavelets (GMWs). The best parameters of generalized Morse wavelets (GMWs) are obtained by using a geological model of the tight reservoir. Second, we extract SSAs of the target reservoir of interest. Then, we perform multi-dimensional data analysis using the principal component analysis (PCA), thus significantly reduce the computational time and storage space for SSAs analysis and visualization. Using red-green-blue (RGB) blending technique we make a high-resolution subsurface depositional facies map from the reduced three principal components from the original multi-dimensional SSAs. Next, we perform unsupervised classification via clustering of SSAs using the fuzzy self-organizing map (FSOM) to generate a seismic facies classification of the reservoir. Finally, we combine multiple linear regression (MLR) and the radial basis function neural network (RBFNN) to provide a quantitative prediction of the reservoir thickness by using preciously drilled wells to train the neural network and to validate the results. Our results illustrate significant variation in reservoir thickness across the field, which can be useful for evaluating reservoir heterogeneity and connectivity. We conclude that our machine-aided multi-dimensional SSAs analysis can be useful for play screening in the reconnaissance phase, prospect generation and maturation in the exploration phase, and well placement in the development phase.