Prediction of Lost Circulation Zones using Support Vector Machine and Radial Basis Function

Drilling deep and high-pressure high-temperature wells have many challenges and problems. One of the most severe, costly and time-consuming problem in the drilling operation is the loss of circulation. The drilling fluid accounts for 25-40% of the total cost of the drilling operation. Any loss of the drilling fluid will increase the total cost of the drilling operation. Uncontrolled lost circulation of the drilling fluid can result in dangerous well control problems and in some cases the loss of the well. In order to avoid loss circulation, many methods were introduced to identify the zones of losses. However, some of these methods are difficult to be applied due to financial issues and lack of technology and the other methods are not accurate in the prediction of the thief zones. The objective of this paper is to predict the lost circulation zones using two different techniques of artificial intelligence (AI). More than 5000 real field data from two wells that contain the real-time surface drilling parameters was used to predict the zones of circulation loss using radial basis function (RBF) and support vector machine (SVM). Six surface drilling parameters were used as an input. The data of the well (A) was divided into training and testing to build the two AI models and then unseen well (B) was used to validate the ability of AI models to predict the zones of lost circulation. The obtained results showed that the two models of AI were able to predict the zones of circulation loss in well (A) with high accuracy in terms of correlation coefficient (R), root mean squared error (RMSE) and confusion matrix. RBF predicted the losses zones with excellent precision of R = 0.981 and RMSE = 0.088. SVM also achieved higher accuracy with a correlation coefficient of 0.997 and root mean squared error of 0.038. Moreover, the RBF model was able to predict the losses zones in the unseen well (B) that was used as a validation for the ability of the AI models with a correlation coefficient of 0.909 and root mean squared error of 1.686.