Prediction of sonic log and correlation of lithology by comparing geophysical well log data using machine learning principles

The well logging technique is used to determine the petrophysical properties like porosity, permeability and fluid saturations of subsurface formations. However, the conventional way of log evaluation is very expensive and tiresome. The post-acquisition processing and inversion provides an alternative to determine the properties of drilled formations. This study proposes a novel approach to predict sonic log, adopting a regression method using a supervised machine learning (ML) algorithm, along with the determination of lithology employing clustering and a neural network approach grounded on the basis of gamma-ray log values and hence creating a correlation between the two. The scarce acoustic data obtained upon the traditional well logging procedure often pose a barrier in further determining the rock physics. Regression analysis, a predictive modeling technique, uses other petrophysical data to predict the sonic wave travel time (shear and compressional) by estimating a relationship between the two variables. The model is trained on a set of 10,000 points with 80% training points giving an 86.314% accuracy result. RMSE and R2 scores for training points and testing points came out to be 2.622 and 0.95, and 2.55 and 0.96, respectively, which helps in the validation of the model. Effective lithology determination is a crucial step of reservoir characterization. Traditional methods of core sample inspection and using well logs, however, cannot meet the needs of real-time due to complex sediment environment and reservoir heterogeneity. To deal with the problem, an unsupervised ML model, K-means clustering, a method of vector quantization grouping unlabeled data into arbitrary clusters based on similarities with respect to distance from the center is used. The model gave the optimum number of clusters as 5 and showed a presence of siltstone, coal and sandstone separated between these clusters. The Silhouette score which tests the accuracy came out to be 0.5840 along with a CH score of 27,192.