Rock typing and causality analysis in unconventional formation using Bayes nets
E. Chekhonin, Alexander Vsevolodovich Nikitin, Dimitri Pissarenko, Yuri A. Popov, Raissa Romushkevich, D.E. Zagranovskaya · Petroleum Research · 2025
Unconventional formations are characterized by high heterogeneity and anisotropy, making it difficult to interpret logging data, perform core analysis, and create accurate petrophysical models. To address this challenge, a new methodology has been developed that uses Bayesian networks to perform rock-typing in unconventional reservoirs. The research combines, for the first time, logging data and thermal core profiling data to enhance classification accuracy based on Bayesian theory. The material used in this study is unique, encompassing extended logging and core data, a large number of full-size core samples, and various Bayesian networks. Constrained-based, score-based, and hybrid algorithms were applied to identify network structures. Comparing different Bayesian networks' performance was made possible by applying the methodology to rock-typing of 2923 full-size samples taken from three wells drilled in the central part of West Siberia through the Bazhenov Formation. Hybrid algorithms − K2-MPC and HC-MWST − gives the networks providing the most successful rock-typing (up to 84% of correctly classified samples). Naive Bayes' rock-typing quality is the worst, which indicates that the input data is not independent. Accounting for some thermal core profiling data improves (up to 6%) the quality of the rock-typing. Lessons we learned in obtaining an effective network structure are described. Bayesian networks and Random Forest provide similar rock typing results, but Bayesian networks give probabilities for samples belonging to each of the rock types studied and also allows specialists to grasp the causal relationships among input parameters. This research offers a new perspective for the analysis of unconventional reservoirs.