Abstract
Pressure prediction has long been one of subject of research focuses in petroleum geology and exploration, but is traditionally limited to moderately overpressured formations due to disequilibrium compaction based on the empirical velocity-pressure model. Overpressures in the Xihu depression are caused by fluid expansion due to the generation of natural gas in source rocks. The resulting abnormally pressured gas reservoirs manifest as a high-pressure structure with the thick mud-coated thin sand (MCS) that remains largely unaddressed in prediction. Such gas-producing overpressures make traditional prediction methods unapplicable because of the significant influence of high gas contents on seismic velocities. We employ a frequency-dependent quality factor Q-pressure petrophysical model because of the sensitivity of acoustic attenuation to overpressures. However, acoustic attenuations from rock physics and sonic logging present an uncertain dependence on pore pressures. This implies that overpressure prediction is a typical physical system that is characterized by both deterministic mechanism and statistical behavior, which resorts to physics-informed deep learning methods. We take the Q-pressure petrophysical model as the activation function of Caianiello convolutional neurons with an attempt to combine deterministic and statistical mechanisms for overpressure prediction. Such convolutional neurons render strong feature extraction and powerful learning ability and are used to construct physics-informed Caianiello convolutional neural networks (CCNNs) for overpressure prediction in the Xihu depression. The comprehensive CCNNs-based seismic inversion scheme for overpressures consists of reservoir petrophysical modeling, deep learning for neural wavelets, well-seismic correlation analysis, deconvolution-based inversion for an initial pressure model, and input-signal reconstruction to improve the initial pressure model. With drill stem tests (DSTs), well logs, and 3D seismic data, we conduct the overpressure prediction associated with cross validation in the Xihu depression, which demonstrates the applicability of the physics-informed CCNNs inversion scheme. Predicting reliability away from the training wells usually depends on the geological complexity of areas under study. Nevertheless, the information representation in the trained CCNNs can be improved gradually by feeding new well data during the development of an oil field.
NSTL