Predicting Formation Tops While Drilling Using Artificial Intelligence

Al-AbdulJabbar, Ahmad (KFUPM) | Elkatatny, Salaheldin (KFUPM) | Mahmoud, Mohamed (KFUPM) | Abdulraheem, Abdulazeez (KFUPM)

OnePetro 

Abstract

Formation tops is one of the important information that is gathered during the exploration and delineation phase. This valuable information aids in setting the casing properly during the development phase and ensure proper zonal isolation between different zones. Every time a new well is drilled, actual formation tops are picked using various methods such as rate of penetration (ROP) charts, gamma ray (GR), formation cuttings and mud logging. These data are used in updating the geological model and in ensuring a proper zonal isolation in critical sections. Each of these methods has its own advantages and limitations such as cost, accuracy, and man power. Most of these methods suffer from a lag in time or depth which prevents the formation tops from being picked instantaneously.

The goal of this paper is to introduce a better method for picking formation tops. It can be a potential alternative to replace other more expensive techniques. The new technique involves the use of drilling mechanical parameters along with the rate of penetration to increase the accuracy of prediction. This will help to detect a true increase or decrease in ROP even if the drilling parameters are fluctuating.

Field data were gathered from two wells with the same bit size and the same formation type. The data were screened and cleaned from any outliers or noise using six different algorithms while retaining the data quality and representation. Using six inputs and four outputs, 30 different sensitivity analyses were conducted including using artificial neural network (ANN) to achieve the best results and prediction accuracy. Well-A was used to train and test the data with 70/30 ratio, while well-B was totally unseen data.

The results obtained showed that ANN can predict formation tops with great accuracy. The best result was found using ANN with 20 neurons and one layer in which correlation of coefficient (R) was 0.94 and 0.98 for both wells. With this new technique, detecting formation changes will be faster compared to other methods since no logs have to be processed and nor any wait is required for cuttings to reach the surface. The formations can therefore be picked in real-time with good accuracy at almost no extra costs because it uses the real-time data which is already available.