Applied Machine Learning for the Imputation of Missing Core Petrophysical Property Data in Clastic and Carbonate Reservoirs

Estimating missing petrophysical data, particularly in permeability/porosity core-analysis measurements is a challenge. The data gaps substantially increase uncertainty in multivariate statistical analysis degrading the accuracy of geomodeling. Data imputation algorithms can be used to estimate such missing values but their performance and reliability vary in providing usable estimates. This study comprehensively compares seven imputation techniques for predicting missing core-measured horizontal and vertical permeability and porosity data in two wells drilled in the North Rumaila oil field a southern Iraqi field; one in the clastic Zubair Formation, and one in the carbonate Yamama Formation. Relative bias (RB) and robustness to outliers were used to assess and compare the estimates of each imputation technique. The results are presented in terms of outlier percentages and displayed using several statistical plots to compare the imputed and original data points. The results reveal that the MICE-CART data imputation method outperforms the other methods with the clastic and carbonate datasets studied. It generates low RB values, and realistic variable density distributions and demonstrates robustness to outliers. The novel workflow (combining relative bias, outlier robustness, and comparison of depth, variable distribution, and box plots) is suitable for application in both clastic and carbonate reservoir formations. It is effective at revealing the relative strengths and weaknesses of specific data imputation methods.