Determination of pore size distribution in tight gas sandstones based on Bayesian regularization neural network with MICP, NMR and petrophysical logs

Abstract

Mercury injection capillary pressure (MICP) is one of the most important means for quantitative study of pore structure, however, its application is often limited due to its expensive, destructive and dangerous characteristics. Nuclear magnetic resonance (NMR) then becomes a common alternative by log-linear conversion of transverse relaxation time (T2) spectrum into pseudo-MICP or pore size distribution (PSD) though the accuracy and generality is unconvincing. Petrophysical logs have rich geological-petrophysical information but are seldom used in pore structure research. In this paper, PSD determination of tight gas sandstones based on Bayesian regularization neural network (BRNN) with MICP, NMR and petrophysical logs is studied. A numerical workflow is developed to preprocess MICP and NMR experimental data and determine PSD. Subsequently, two BRNN models are constructed respectively with joint laboratory T2 spectrum and petrophysical logs as main inputs and pore radius and mercury saturation from MICP measurement as training target. It is revealed that, accurate PSD can be determined from prediction results of BRNN models, wherein the NMR-based BRNN can be employed to obtain PSD from more easily accessible NMR measurement and the log-based one is expected to be generalized in the field. Conversion of PSD by fitting pore radius with T2 through their piecewise linear or quintic polynomial relation in log-log coordinate is equivalent to underfitting or overfitting in accuracy of machine learning and is confined to laboratory NMR with joint MICP. Detailed discussion on BRNN modeling, pore radius fitting and the corresponding numerical workflow proposed here may provide valuable insight into petrophysical-log analysis and quantitative evaluation on pore structure of complex reservoirs.