Porosity prediction using semi-supervised learning with biased well log data for improving estimation accuracy and reducing prediction uncertainty

Porosity characterization is of profound significance for seismic inversion and hydrocarbon prediction. Although semi-supervised learning (SSL) based methods have been used to boost prediction accuracy and lateral continuity of supervised learning (SL) inverted subsurface properties, their variations are relatively limited since the relationships between the data and the parameter model are straightforward in most reported cases. To further figure out their essential differences, we proposed the SSL-based network (SSLBN) for reservoir porosity prediction using seismic and well log data with disparate complexity and quality, and compared it with the SL-based network (SLBN). The SSLBN comprises a data-driven inverse model named decoder and a data-driven forward model named encoder based on the bidirectional-gated recurrent units. The architecture of the SLBN is the same as the encoder. Trained by several seismic-to-well pairs and numerous unlabelled seismic logs, the SSLBN learns the physical process from input single-trace observed seismic log to the intermediate porosity log, and the inverted porosity to the output generated seismic log. We first prepare the porosity model with biased or unbiased labels, the convolution model (CM) and reverse time migration (RTM) based synthetic seismic data, and then implement SL- and SSL-based statistical tests. The synthetic data examples demonstrate that the SSLBN has significant preponderance over the SLBN in the scenes of the RTM imaged seismic data and biased porosity labels. Compared with the SLBN, the physical regularization of the data misfit in the SSLBN improves estimation accuracy and reduces prediction uncertainty of porosity. Finally, statistical tests on a braided river deposited field data example illustrate that the SSLBN can generate more geologically trustworthy porosity models and indicate the oil layers of high porosity sandstone reservoirs.

Automated summary

Porosity characterization is important for seismic inversion and hydrocarbon prediction. We proposed a SSL-based network (SSLBN) for reservoir porosity prediction, and compared it with a SL-based network (SLBN). The SSLBN, trained with seismic-to-well pairs and unlabelled seismic logs, learns the physical process from observed seismic log to inverted porosity and generated seismic log. Results show that the SSLBN outperforms the SLBN in scenes of RTM imaged seismic data and biased porosity labels, improving estimation accuracy and reducing prediction uncertainty.