Porosity Prediction from Seismic Using Machine Learning: Example from North-West Shelf Offshore Australia

Porosity has always been an essential property for determining reservoirs’ volumetric; however, determining porosity with an acceptable range of certainty carries many challenges. The deep-water Plover reservoir in Poseidon area, North Western Australia exhibits a special depositional environment, unique interior structure, and a complex porosity distribution that make predicting reliable reservoir properties challenging. This study aims to enhance the characterization of the Poseidon reservoir in terms of porosity using a combination of reservoir geophysics (seismic attributes and petrophysics) through machine learning (ML) techniques. Three methods of porosity estimation from seismic data have been implemented and compared: (1) multi-linear regression (MLR), (2) probabilistic neural network (PNN), and (3) deep feed-forward neural networks (DFNN). The seismic data available is a post-stack volume inverted to derive P-impedance. After the inversion, training data at well locations were analyzed, and statistical relationships were established between the porosity log, the seismic data, and the seismically derived P-impedance. Cross-validation was used to assess the reliability of the derived relationships. The probabilistic neural network (PNN) showed promising results far better than other comparative methods. Apart from PNN, the deep feed-forward neural network (DFNN) was also tested but gave limited success due to the scarcity of labeled data. The lack of labeled data has limited the optimum prediction of subsurface properties to a large extent. The predicted porosity from PNN has revealed geological features that otherwise are not seen in simple seismic inversions or seismic attribute analysis.