Modeling and Inversion for Wave Propagation in Tight Sandstone Reservoirs Using Machine Learning

Abstract Understanding wave propagation in subsurface reservoirs is an important topic in exploration geophysics. Using machine learning (ML), this study aims to develop a hybrid modeling approach that uses data techniques while maintaining the reliability of poroelasticity theory. Simplified dynamic equations for seismic propagation in sandstone reservoirs are established in two steps: Biot‐Rayleigh theory is established and then an optimization algorithm in ML is used to identify a simplified equation and calculate a local fluid flow term, which is responsible for wave attenuation, and some of the elastic constants and factors such as the volume ratio of inclusions. The effectiveness of the approach is first tested on synthetic data, and it is shown that almost the same dispersion and attenuation as the original model can be predicted. Data from experimental and borehole measurements are then considered. Examples show that with a few data points the wave velocity can be accurately predicted in different frequency ranges. Although the model has a certain extrapolation capability, the coverage of training data is still required. Finally, the approach is extended to perform porosity inversion. The proposed technique can be extended to reservoirs with different lithologies.