Reconstructing Missing and Non-Recorded Well Log Data Using Machine Learning Techniques.

Abstract Petrophysical estimation of rocks are carried out for lithologic characterization and to quantify the reservoir hydrocarbon prospects. During drilling operation, the continuous recording of well log data is performed using Logging Whilst Drilling (LWD) tools and sometimes the LWD tool failure results in loss of data especially when decision was made to drill ahead due to project economics resulting in gaps or unrecorded data. This study is aimed at using Machine Learning (ML) techniques in predicting missing and unrecorded data using offsets well data. During the study detailed literature review addressing current industry trends in the handling gaps in petrophysical log were reviewed, then four ML models were trained and used in predicting lost data entries using offsets well log data. The models utilized comprised CatBoost (CB), Gradient boost (GB), Random Forest (RF) and Support Vector Machine (SVR). The ML algorithms were first trained with 21595 data points using a 70:30 percentage ratio for training and testing respectively. After that, 8650 data points from a different well were used to validate the developed model. The real-time drilling parameters required as input data includes Gamma ray(GR), neutron (NPHI), bulk density (RHOB), photoelectric absorption factor (PEF), the borehole diameter, (CALI). The modeling result showed a good match between the actual and predicted (logs) with coefficients of determination for training and testing dataset of 0.99 and 0.97 for CatBoost and RF respectively which are the two best predictors. The main added value of this approach is that the input drilling parameters are readily available in real-time and can timely aid in well placement decisions which improves the productivity of the well.