Semi-supervised seismic data and well logs integration for reservoir lateral porosity prediction

Summary Traditional porosity prediction methods usually adopt two consecutive steps including seismic inversion and petrophysical modelling to convert seismic data into porosity. Machine learning can take full advantage of available geophysical information to directly build the nonlinear mapping for predicting porosity from seismic data. To realize the one-step reservoir porosity estimation, we propose the semi-supervised recurrent neural networks (SSRNNs) based porosity modelling method. SSRNNs include an encoder subnet and a decoder subnet. The encoder simulates the generalized seismic inversion to convert the input post-stack seismic data into the predicted porosity, and the decoder acts as a forward model to make the predicted porosity can return to the generated seismic data and reduce resolution space. In addition, seismic data at the non-well positions are randomly selected in each iteration of SSRNNs to boost the lateral continuity of the predicted porosity result. Without the demand of some approximate assumptions and accurate elastic parameters, well logs and seismic data at well locations and non-well locations are integrated into SSRNNs to directly predict high-precision porosity from seismic data. A numerical model example and a real data example are used to verify the effectiveness and accuracy of the SSRNNs based reservoir lateral porosity prediction method.