Predicting hydrogen and oxygen indices (HI, OI) from conventional well logs using a Random Forest machine learning algorithm

Hydrogen Index (HI) and Oxygen Index (OI) are two critical parameters for assessing the hydrocarbon potential and depositional environment of any source rocks. The most common method to measure these values is to use programmed pyrolysis on drill samples. However, this method can be very time consuming, expensive, and in many cases much of the well bore may be overlooked due to biased sampling. Geochemical parameter predictions from wireline logs (i.e., Passey) have been used in the past to varying success. This is largely because petrophysical predictions often attempt to solve for linear regression solutions where this may not be the case. Here we evaluate the use of a Random Forest (RF) machine learning (ML) model to predict HI and OI from four wells from the offshore east coast of Newfoundland, Canada. The model was trained and tested using programmed pyrolysis data, organic petrology techniques, and wireline logs for prediction. The model was evaluated using mean absolute error (MAE), root mean square error (RMSE), correlation of determination (R2), and Spearman's rank correlation (R2). Excellent correlation coefficients were observed for RF model predictions for HI and OI that range 0.90 to 0.98 and 0.90 to 0.95 R2 respectively. The MAE for HI and OI values range 17.30 to 52.48 and 2.82 to 12.79 respectively. The RMSE for HI and OI range 21.43 to 71.51 and 3.85 to 16.82 respectively. The Spearman's rank correlation for HI and OI range 0.87 to 0.97 and 0.90 to 0.96 respectively. This study confirms that the use of ML models can be extremely useful to predict geochemical parameter from wireline logs.

Automated summary

This study evaluated the use of Random Forest (RF) machine learning model to predict Hydrogen Index (HI) and Oxygen Index (OI) from four wells offshore east coast of Newfoundland, Canada. The model showed excellent correlation coefficients and low error ranges for predicting HI and OI, demonstrating the great potential of machine learning models in predicting geochemical parameters from wireline logs.