Severe fluid losses while drilling carbonate reservoirs have considerably increased well construction time and costs. Such extra expenses are mainly related to wasted time while struggling against such losses, material costs and its delivery/availability logistics. Besides the economic impact, severe circulation losses have HSE (Health, Safety & Environment) impacts since there is a risk of losing well control when losses can evolve to hydrocarbon inflow and simultaneous loss and gain issues. Extreme situations may lead to temporary or even definitive well abandonment.
Fluid loss predictions are usually performed by a specialist with knowledge of the geological model and the drilling history of the field. Such approach has proved to be a hard task, with limited success, especially in the pre-salt carbonate reservoirs due to their high structural and facies heterogeneity.
This study aims to improve prediction (as compared to the conventional expert-based approach) of the geological structures that might lead to severe fluid losses with impact on well construction costs by focusing on uncertainty reduction of critical resources allocation, such as Managed Pressure Drilling (MPD) and loss control materials (LCM).
Artificial Intelligence (AI) techniques have proved to be useful with high success rate in complex problem solving in many industrial segments. The focus of the present study was to search for AI algorithms to correlate seismic attributes, well logs, fluid loss occurrences and information from geologic and reservoir flow models. A pilot area comprising 38 wells drilled in Santos Basin (Brazil) was chosen for the present analysis. The first step was to use this data set to map the search space of the algorithms, i.e., to identify the critical intervals for severe losses. Information gain tests related to the fluid loss rate (dependent variable) were performed aiming to identify the most relevant independent variables for the case of severe losses prediction and to discard the ones with minor contribution.
Among the tested classifiers, an ensemble of Naive Bayes & Perceptron Neural Network had the best performance at predicting severe fluid losses for the pilot area. A global hit rate of 84% was achieved for metrics evaluated under a well-based standpoint. Blind test with 11 wells (from a different set) returned 82% of global hit rate. These results are considered superior to the ones obtained through the conventional approach. It is important to mention that due to the great uncertainty of the related variables, the output cannot be more accurate than the precision of the original data employed.
These results show a great potential for the use of AI techniques on severe fluid losses prediction in pre-salt carbonates. Therefore, the AI approach is being incorporated as a new tool to support the field experts to improve the performance of the predictions.