A Novel Data Assimilation Methodology for Predicting Lithology Based on Sequence Labeling Algorithms

Abstract:

A hidden Markov model (HMM) and a conditional random fields (CRFs) model for lithological predictions based on multiple geophysical well-logging data are derived for dealing with directional non-stationarity through bi-directional training and conditioning. The developed models were benchmarked against their conventional counterparts, and hypothetical boreholes with the corresponding synthetic geophysical data including artificial errors were employed. In the three test scenarios devised, the average fitness and unfitness values of the developed CRFs model and HMM are 0.84 and 0.071, and 0.81 and 0.084, respectively, while those of the conventional CRFs model and HMM are 0.78 and 0.091, and 0.77 and 0.099, respectively. Comparisons of their predictabilities show that the models designed for directional non-stationarity clearly perform better than the conventional models for all tested examples. Among them, the developed linear-chain CRFs model showed the best or close to the best performance with high predictability and a low training data requirement.

 

Keywords: one-dimensional lithological characterization, sequence labeling algorithm, conditional random fields, hidden Markov model, borehole, geophysical well-logging data.