Deep learning a poroelastic rock-physics model for pressure and saturation discrimination

Abstract

Determining saturation and pore pressure is relevant for hydrocarbon production as well as natural gas and [Formula: see text] storage. In this context, seismic methods provide spatially distributed data used to determine gas and fluid migration. A method is developed that allows the determination of saturation and reservoir pressure from seismic data, more accurately from the rock-physics attributes of velocity, attenuation, and density. Two rock-physics models based on Hertz-Mindlin-Gassmann and Biot-Gassmann are developed. Both generate poroelastic attributes from pore pressure, gas saturation, and other rock-physics parameters. The rock-physics models are inverted with deep neural networks to derive saturation, pore pressure, and porosity from rock-physics attributes. The method is demonstrated with a 65 m deep unconsolidated high-porosity reservoir at the Svelvik ridge, Norway. Tests for the most suitable structure of the neural network are carried out. Saturation and pressure can be meaningfully determined under the condition of a gas-free baseline with known pressure and data from an accurate seismic campaign, preferably cross-well seismic. Including seismic attenuation increases the accuracy. Although training requires hours, predictions can be made in only a few seconds, allowing for rapid interpretation of seismic results.