Correlation of artificial neural network and multi-attribute transformation: a test case to predict the effective porosity of Cretaceous sandstones of the Sembar Formation, southeast Pakistan

Accurate prediction of reservoir properties through linear transformation and regression methods are successful in limited cases but are often geologically unrealistic and have no concrete theoretical foundation. Artificial Neural Network’s (ANN’s) have emerged as an effective tool for deriving nonlinear mathematical relationships between seismic attributes and well logs that are theoretically plausible and may prove geologically realistic. In this paper, we devise a methodology to integrate rock physics analysis, seismic inversion, multi-attribute transformation, and Feedforward Neural Network (FNN) modeling for accurate inter-well reservoir property predictions. We test this methodology on well logs and seismic data from the Cretaceous sandstones of the Sembar Formation, Southern Indus Basin, Pakistan. Viable productive gas zones are identified through rock physics and Model Based Inversion (MBI) analyses. Five volume-based seismic attributes are sequentially calculated through forward stepwise regression and cross-validated for inter-well porosity prediction. When a Probabilistic Neural Network (PNN) is trained in a non-linear mode integrated with multi-attribute transformation, correlation ( r 2 ) is improved from 72% to 88% between seismic attributes and porosity derived from logs. The PNN-derived porosity distribution is geologically more realistic than linear transformation and regression methods, supporting our model’s validity. We suggest that it is theoretically possible for the ANN to make predictions about any attribute of the reservoir via bridging target logs and seismic data within a short computation time.