Using Semi-Supervised Convolutional Neural Networks for Porosity Modeling Over a Fluvio-Deltaic Triassic Gas Field

Robust estimation of rock properties, such as porosity and density, from geophysical data, i.e. seismic and well logs, is essential in the process of subsurface modeling and reservoir engineering workflows. Such properties are accurately measured in a well; however, due to high cost of drilling, such direct measurements are limited in amount and sparse in space within a study area. On the contrary, 3D seismic data illuminates the subsurface of the study area throughoutly by seismic wave propagation; however, the connection between seismic signals and rock properties is implicit and unknown, causing rock property estimation from seismic only to be a challenging task and of high uncertainty. An integration of 3D seismic with sparse wells is expected to eliminate such uncertainty and improve the accuracy of static reservoir property estimation. This paper investigates the application of a semi-supervised learning workflow to estimate porosity from a 3D seismic survey and 36 wells over a fluvio-deltaic Triasic gas field. The workflow is performed in various scenarios, including purely from seismic amplitude, incorporating a rough 6-layer deposition model as a constraint, and training with varying numbers of wells. Good match is observed between the machine prediction and the well logs, which verifies the capability of such semi-supervised learning in providing reliable seismic-well integration and delivering robust porosity modeling. It is concluded that the use of available additional information helps effectively constrain the learning process and thus leads to significantly improved lateral continuity and reduced artifacts in the machine learning prediction. The semi-supervised learning can be readily extended for estimating more properties and allows nearly one- click solution to obtain 3D rock property distribution in a study area where seismic and well data is available.