Sonic logs are required for well-seismic ties, an essential step in the exploration workflow. However, due to various constraints, sonic logs might not be recorded in several wells within a field. This is especially true for older wells. It is also seen that basic logs like GR, RT, RHOB, NPHI are very likely to be recorded in all wells.
So far, geoscientists approximate missing sonic logs from sonic logs of nearby wells using empirical techniques. Such techniques might not be a robust representation of the the sub surface environment. Further, these techniques are manual and inconsistent. Moreover, such missing logs are approximated as required by a certain study and may not be shared for general consumption by all projects.
I propose a new data driven approach to predict missing sonic logs on a field scale. I hope to predict sonic logs for all the wells in a field and make them easily available.
Figure 1 : Solution Approach
Briefly, we train a field specific machine learning ( ML ) model on wells which have both basic logs ( GR, NPHI, RHOB, RT) and sonic logs (DTC, DTS). This ML model is used to predict missing sonic logs in wells of the same field using their respective basic logs.
Well Logs , stored on EPINET are already organised based on fields and wells and stored in LAS format.
Within each field, LAS files could be automatically read to determine the curves ( logs ) present per well. For this solution, the available data could be divided into two parts.
The objective is to build an ML model that can predict sonic logs using basic logs.
Wells with Basic logs ( Prediction Input ) and sonic Logs ( Prediction Target ) are split into training and validation datasets. The training dataset is processed by an ML algorithm to build an ML model. Various algorithms have been shown to be effective ( Random Forest, Gradient Boosting, Artificial Neural Networks ).
The effectiveness of the trained ML model in predicting sonic logs is assessed ( Model Validation ) using the validation dataset. For model validation, the trained ML model is used to predict sonic logs from basic logs in the validation dataset. The predicted sonic log is compared with true sonic logs in the validation dataset to determine the accuracy / error rate of the ML model. A good ML model performs similarly on training and validation dataset. Performance is measured using Mean Squared Error and coefficient of determination.
Such trained and validated ML model could be stored on EPINET.
Thereafter, missing sonic logs can be predicted using stored ML model using basic logs as input. Such basic logs + predicted sonic logs are written to new LAS files that can be readily consumed by all related geoscience workflows.
ML models trained on well logs in a particular field have been shown to be effective in predicting missing sonic logs in wells throughout the same field.