Mature field operators collect log data for tens of years. Collection of log dataset include various generation and multiple vintages of logging tool from multiple vendors. Standard approach is to correct the logs for various artefacts and normalize the logs over a field scale. Manually conducting this routine is time consuming and subjective. The objective of the study was to create a machine learning (ML) assisted tool for logs in a giant Lower Cretaceous Carbonate Onshore field in Abu Dhabi, UAE to automatically perform data QC, bad data identification and log reconstruction (correcting for borehole effects, filling gaps, cleaning spikes, etc.) of Quad Combo well logs.
The study targets Quad Combo logs acquired since mid-60's. Machine learning algorithm was trained on 50 vertical wells, spread throughout the structure of the field.
The workflow solution consists of several advanced algorithms guided by domain knowledge and physics based well logs correlation, all embedded in an ML-data-driven environment. The methodology consists of the following steps:
Outliers detection and complete data clustering.
Supervised ML to map outliers to clusters.
Random Forest based ML training by clusters, by logs combination on complete data.
Saved models are applied back to the whole data including outliers and sections with one or several logs missing.
Validation and Blind test of results.
Models can be stored and re-used for prediction on new data.
The ML tool demonstrated its effectiveness while correcting logs for outliers’ like Depth Offsets between logs, identifying Erroneous readings, logs prediction for absent data and Synthetic logs corrections. The tool has a tendency to harmonize logs. First test demonstrated robustness of the selected algorithm for outliers’ detection. It cleaned data from most of contamination, while keeping good but statistically underrepresented logs readings.
Clustering algorithm was enhanced to supplement cluster assignment by extraction of the corresponding probabilities that were used as a cut-off value and utilized for a mixture of different ML models results. This application made results more realistic in the intervals where clustering was problematic and at the transition between different clusters.
Several intervals of bad and depth shifted logs corrections were noticed. Outliers’ corrections for these logs was performed the way that at Neutron-Density or Neutron-Sonic cross-plots points were moved towards expected lithology lines. Algorithm could pick-up hidden outliers (such as synthetic logs) and edited the logs to make it look intuitively natural to a human analyst.
The work successfully demonstrated effectiveness of ML tool for log editing in a complex environment working on a big dataset that was subject of manual editing and has number of hidden outliers. This strong log quality assurance further assisted in building Rock Typing based Static Model in complex and diagenetically altered Carbonates.