Development and evaluation of an effective machine learning model for well log prediction: A case study of sonic log prediction of zircon field Niger-Delta Nigeria

Odiegwu, Chukwukelu Johnpaul; Chinwuko, Ifeanyi Augustine; Aniwetalu, Emmanuel Udenwa; Umeh, Maureen Chioma; Igwebudu, Callista Nwanneka; Akpudo, Chinaza Janefrances

doi:10.21608/jaiep.2025.430788.1027

Development and evaluation of an effective machine learning model for well log prediction: A case study of sonic log prediction of zircon field Niger-Delta Nigeria

Document Type : Original Article

Authors

Department of Applied Geophysics, Nnamdi Azikiwe University, Awka 5025, Nigeria

10.21608/jaiep.2025.430788.1027

Abstract

The accurate prediction of sonic log data is critical for subsurface characterization and reservoir management in hydrocarbon exploration. Conventional methods of predicting missing well logs which often relied on interpolation techniques or empirical correlations are limited in their ability to capture the complex, nonlinear relationships that exist in subsurface formations. In this study we present a methodology for predicting a missing log. Three wells from the Zircon field in the Niger-Delta were used in the study: Well 7, 1, and 6 for training, validation and prediction phases respectively. The results of the preprocessing steps which involved outlier removal, missing value handling and filtering operation with Butterworth lowpass filter, were effective in improving the correlation among predictor variables and target Sonic log. Five models were initially used in the training and validation phases using the Sci-kit ML module in Python. The RF model was finally selected for the prediction phase having out-performed other models with a RMSE, MEDAE and R-SQUARED SCORE values of 2.5886µs/ft, 1.0380µs/ft, 0.9642 in the testing phase and 6.5588µs/ft, 3.6356µs/ft and 0.7695 in the validation phase respectively. A supplementary qualitative well correlation analysis performed using the training, validation and prediction wells gave satisfactory results based on similarities in the sonic log character and trend. The qualitative well correlation provided a crucial geological validation of the model's output. The findings of this research could significantly reduce the need for extensive well logging operations and provide a framework for integrating machine learning techniques into petroleum geoscience.

Keywords