Machine Learning for Sonic Logs Prediction: A Case Study from the Niger Delta Basin in the Gulf of Guinea

Abstract

Abstract Sonic logs are very essential for rock type identification, hydrocarbon typing, rock physics modelling, and reservoir characterization. However, they are seldom available due to high costs of acquisition or measurement errors. Empirical formulas and petroelastic models that are often used to predict missing sonic log data may not produce accurate velocity profiles and are limited to specific geologic settings. Using sonic logs from wells in the CONDA field located in the deep offshore Niger delta basin of the Gulf of Guinea, we demonstrate that machine learning algorithms can be used to predict sonic log data if suitable quantitative relationships exists between it and available well logs. Preprocessing such as outlier removal, missing data filling and normalization was applied to the well logs before using them as training datasets for the model prior to applying several machine learning algorithms to build a predictor model for missing DTP and DTS sonic logs. The results of the training using several machine learning algorithms showed that the Gradient Boost Regressor (GBRT) Algorithm was more robust based on higher accuracy and lower root-mean-squared errors (RMSE). Validation of the prediction model at blind wells was quite good, with coefficient of determination or goodness-of-fit (R2) scores of 0.88 to 0.99 and generally low root mean square errors (RMSE). QC of the predictive model performed using qualitative well correlation analysis between a well with actual DTP and DTS sonic logs and another with predicted DTP and DTS sonic logs gave very satisfactory results based on similarities in log character and trend. The results of our study show that in comparison to sonic log prediction using empirical formulas and/or petroelastic models which is fraught with limitations, machine learning can be used as a robust alternative.