PREDICTION OF ROCK AND FLUID PROPERTIES USING MACHINE LEARNING, MULTI-SCALE SHORT TIME FOURIER TRANSFORM, AND ROCK PHYSICS

Direct hydrocarbon indication using elastic seismic inversion can be optimized using a weighted difference between compressional-wave and shear-wave impedances. Application to a wide variety of well log and laboratory measurements suggests that the empirically optimized weighting may differ from direct theoretical calculations made using Gassmann’s equations. Combining laboratory and log measurements for sandstones having a broad range of frame moduli reveals a simple linear empirical relationship between the optimized fluid discrimination weights and the squared velocity ratio of brine-saturated sandstones. For the Volve oil field in the North Sea, well log velocity prediction using machine learning methods achieves significantly better prediction error than does a representative, but highly simplified, rock physics model. A hybrid approach, combining machine learning and rock physics modeling, improves prediction robustness, with the average coefficient of determination (R2) score increasing by 13.3% for P-wave velocity prediction relative to the rock physics modeling and 5.0% for P-wave velocity prediction relative to the machine learning results. In predicting well log properties from seismic data, the scale differences between geological layering and seismic resolution can be addressed utilizing spectral decomposition accomplished with a multi-scale Fourier transform combined with machine learning methods. For synthetic data, as compared to conventional machine learning methods, the correlation coefficient in porosity prediction increased from 0.74 to 0.94. For real data from the Volve field in the North Sea, the correlation coefficient increased from 0.41 to 0.81. Machine learning predictions of absolute rock properties such as seismic impedance, are limited by the lack of low-frequency information in the seismic trace. This can be addressed by incorporating various seismic/geological attributes that contain abundant low frequency components, such as relative geological age and interval velocity. These can be used to predict the low frequency impedance trend below the seismic bandwidth from the seismic signal bandwidth using deep neural networks. The results from a real data case of the Midland basin reveal that this approach improves the prediction accuracy of the low frequency components of acoustic impedance, with the best improvement of 57.7% compared with the results predicted using the well log interpolation method.