Interpretation of subsurface lithology has been an integral part of bottom-hole investigation while drilling a well for hydrocarbon exploration. Conventional methods such as mud logging and open-hole logging have proven to be effective in this aspect. But there arises a need of the real time interpretation of subsurface drilling lithologies to have a better control over desired well path and economics, which is very limited through available conventional methods. This paper aims to provide a classifier that can deliver a justified chance of success in lithology estimation based on input matrix provided to the classifier. The dataset is taken from open source platform (Equinor), which provided the drilling history of 15_9_F_15_D well of North Sea Basin along with the interpreted lithology. This dataset is used to prepare the input matrix and the corresponding output used for the classifier. It incorporates seven drilling parameters, i.e. rate of penetration (ROP), weight on bit (WOB), bit rotation per minute, total bit rotations, corrected drilling exponent, mud flow & mud weight and the corresponding recorded subsurface lithology for each foot. It is made sure that the chosen drilling parameters have a direct impact on the rate of penetration with changing physical characteristics of the formations being penetrated. The formations are calibrated with the integer values from 1 to 5 for easier numerical computations. The prepared dataset is divided into two components i.e. training Set (80%), and test Set (20%) with feature scaling implemented. The most intuitive classification model among all the tested classification algorithms i.e. multiple classification algorithm, artificial neural network, and KNN search method is stated based on their individual F (1) score on the test set. It is observed that the cost function for the multiple classification model stopped reducing after the first few iterations irrespective of learning rate value alterations, which induced a limit on the accuracy provided by the chosen algorithm. It is observed that the performance of ANN is quite good for the lithology present in majority, such as claystone. However, the performance degrades exponentially for the lithologies occurring in minorities such as sandstone, dolomite, marl and limestone. The results obtained from KNN classification model not only provides an accuracy of 95% in claystone, 92% in sandstone and 89% in marl which makes this model the most appropriate classifier to be used among the three.
The preliminary results, through the comparison among the three classifiers not only provides the most intuitive model for subsurface lithology prediction using real-time drilling data, but also gives an idea about the most appropriate classifier to be used in the future reference, be it any basin, given that the same set of drilling parameters are used as the input matrix for training the classifier.